ADVANCES I N

CLIMBING AND
WALKING ROBOTS
This page intentionally left blank
 ADVANCES IN
CLIMBING AND
WALKING ROBOTS
Proceedings of 10th International Conference (CLAWAR 2007)
Singapore 16 - 18 July 2007

editors
Ming Xie
Nanyang Technological University, Singapore

Steven Dubowsky
Massachusetts Institute of Technology, USA

Jean-Gu y Fon tai ne

Massachusetts lnstitute of Technology, USA

M Osman Tokhi
University of Sheffield, UK

Gurvinder S Virk
University of Sheffield, UK

1; World Scientific
N E W JERSEY * LONDON * SINGAPORE * BElJlNG * SHANGHAI * HONG KONG * TAIPEI - CHENNAI
Published by
World Scientific Publishing Co. Pte.Ltd.
5 Toh Tuck Link, Singapore 596224
USA ofice: 27 Warren Sbeet, Suite 401-402, Hackensack, NJ 07601
UK ofice: 57 Shelton Street, Covent Garden, London WC2H 9HE

British Library Cataloguing-in-PublicationData

A catalogue record for this book is available from the British Library.

ADVANCES IN CLIMBING AND WALKING ROBOTS

Proceedings of 10th International Conference (CLAWAR 2007)
Copyright 02007 by World Scientific Publishing Co. Re. Ltd.
All rights reserved. This book, or parts thereoJ; may not be reproduced in any form or by any means,
electronic or mechanical, including photocopying, recording or any informationstorage and retrieval
system now known or to be invented, without written permission from the Publisher.

For photocopying of material in this volume, please pay a copying fee through the Copyright
Clearance Center, Inc., 222 Rosewood Drive, Danvers, MA 01923, USA. In this case permission to
photocopy is not required from the publisher.

ISBN-13 978-981-270-815-1
ISBN-10 981-270-815-4

Printed in Singapore by World Scientific Printers (S) Pte Ltd

 PREFACE

Robotics is an exciting field in engineering and natural sciences. Robotics

has already made important widespread contributions and impact in industrial
robots for tasks such as assembly, welding, painting, and material handling.
In parallel, we have also witnessed the emergence of special robots, which
perform valuable tasks, in non-industrial environments such as in search and
rescue, de-mining, surveillance, exploration, and security missions. Furthermore,
research and development works are currently in progress in the robotics
technology for use in the domestic and professional service sector. The emergence
of mobile machines, such as climbing and walking robots, for these missions in
un-structured environments, has significantly broadened challenges that must
be considered by robotics research. This includes not only the technological and
engineering aspects including standardisation, but also socio-economic and
ethical aspects.
CLAWAR 2007 is the tenth in a series of international conferences
organised annually since 1998 with the aim to report on latest research and
development findings and to provide a forum for scientific discussion and debate
within the mobile service robotics community. The series has grown in its
popularity significantly over the years and has attracted researchers and developers
from across the globe. The CLAWAR 2007 proceeding reports on the latest
scientific and developmental achievements, future challenges and exciting
applications of mobile machines in general, in particular climbing and walking
robots. Eighty-five technical presentations by authors from 22 countries covering
five continents were presented at the CLAWAR 2007 conference, held in
Singapore between 16-18 July 2007. The text of the proceedings is organized into
five parts: Plenary Introduction, Advances in Climbing Robots, Advances in
Walking Robots, Advances in Humanoid Soccer Playing Robots, and Supporting
Technologies.
The editors would like to thank members of the International Programme
Committee, International Technical Advisory/Organising Committee and National
Organising Committee for their hard work in creating a well-run and productive
meeting and for their efforts in reviewing the submissions. We would like to

V
vi

thank the authors for their quick response to comments and suggestions of the
reviewers. It is hoped that this edition of the CLAWAR conference proceedings
will form a valuable addition to the scientific and developmental knowledge in
mobile robotics.

M. Xie
Nanyang Technological University, Singapore
General Chair

S . Dubowsky
Massachusetts Institute of Technology, USA
General Co-Chair

J. G. Fontaine
Italian Institute of Technology, Italy
General Co-Chair

M. 0.Tokhi
University of Shefield, U.K.
Programme Chair

G. S . Virk
University of k e d s , U.K.
Programme Co-Chair

15 May 2007
 CONFERENCE COMMITTEES

Honorary Advisor, General Chairs and Programme Chairs

for the Tenth International Conference on Climbing and Walking Robots

K. Y . Lam (Honorary Advisor) - NTU, Singapore

M. Xie (General Chair) - NTU, Singapore
S. Dubowsky (General Co-Chair) - MIT, USA

J. G. Fontaine (General Co-Chair) - IIT, Italy

M. 0. Tokhi (Program Chair) - University of Sheffield, U.K.

G. S. Virk (Program Co-Chair) - University of Leeds, U.K.

INTERNATIONAL ADVISORY COMMITTEE

for the Tenth International Conference on Climbing and Walking Robots

Manuel Armada - Spain

Yvan Baudoin - Belgium

Kasten Berns - Germany

Philippe Bidaud - France

Bryan Bridge - U.K.
Chan-Hin Kam - Singapore

Krzysztof Kozlowski - Poland

Giovanni Muscat0 - Italy

Lakmal Seneviratne - U.K.

vii
 ...
Vlll

INTERNATIONAL PROGRAMME COMMITTEE

for the Tenth International Conference on Climbing and Walking Robots

A K M Azad - USA A T de Almeida - Portugal

K Althoefer - UK R Arkin - USA

C Balaguer - Spain D P Barnes - UK

J Billingsley - Australia C Bostater - USA

K Bouazza-Marouf - UK M Buehler - USA

G Bugmann - UK S Cameron - UK

C Chevallerau - France H Cruse - Germany

S Cubero - Australia J S Dai - UK

B Davies - UK R Dillmann - Germany

L Fortuna - Italy T F'ukuda - Japan

P Gonzlez de Santos - Spain S S Ge - Singapore

V Gradetsky - Russia A Halme - Finland

R Hasselvander - France E Herrera - Colombia

N Heyes - UK S H'irose - Japan

M A Hossain - UK D Howard - UK
0 Kaynak - Turkey P Kiriazov - Bulgaria
P Kopacek - Austria K Kwok - Singapore

D Lefeber - Belgium J Lopez-Coronado - Spain

C Melhuish - UK R Molfino - Italy

N K MSirdi - France D E Okhotsimsky - Russia

E Dupuis - Canada H R Pota - Australia

R Quinn - USA M Ribeiro - Portugal
J Sa da Costa - Portugal M A Salichs - Spain

T P Sattar - UK M H Shaheed - UK
G Seet - Singapore L Steinicke - Belgium

A Vitko - Slovakia K J Waldron - USA

R Walker - UK H Worn - Germany
A Yigit - Kuwait A Zomaya - Australia
 ix

NATIONAL ORGANISING COMMITTEE

for the Tenth International Conference on Climbing and Walking Robots

Z. W. Zhong (Publicity) - NTU, Singapore

K. H. Low (Workshops) - NTU, Singapore
C. J. Zhou (Exhibitions) - SP, Singapore
C. M. Chew (Technical Visits) - NUS, Singapore

D. Hsu (Technical Visits) - NUS, Singapore

A. P. New (Awards) - DSO, Singapore

H. Y. Yu (Awards) - DSO, Singapore

SECRETARIAT
for the Tenth International Conference on Climbing and Waking Robots

Shirley Soh - CCE, NTU, Singapore

This page intentionally left blank
 CONTENTS

PREFACE V

CONFERENCE COMMITTEES vii

Plenary Introduction

BIPEDAL HUMANOID ROBOT AND ITS APPLICATIONS 3

Astuo Takanishi

CLIMBING UP THE WALL 5

John Billingsley

FROM MICRO TO NAN0 AND SWARM ROBOTS 15

Heinz Worn, Ramon Estaiia, Heiko Hamaun and Marc Szymanski

CLIMBING ROBOTS FOR NONDESTRUCTIVE TESTING: 25

HISTORICAL PERSPECTIVE AND FUTURE TRENDS
Bryan Bridge

BIOMECHANICS AND ROBOTICS 33

Neville Hogan

Advances in Climbing Robots

A CPG WITH FORCE FEEDBACK FOR A STATICALLY STABLE 31

QUADRUPED GAIT
Jose' Cappelletto, Pablo Estkvez, Gerard0 Fernandez-Lopez and
Juan Carlos Grieco

A SLIDING SOCK LOCOMOTION MODULE FOR A RESCUE ROBOT 47

Luca Rimassa, Matte0 Zoppi and Rezia Moljino

A WHEELED WALL-CLIMBING ROBOT WITH A CLIMBING LEG 55

Yili Fu, Zhihai Li, Hejin Yang and Shuguo Wang

AN EVOLVED NEURAL NETWORK FOR FAST QUADRUPEDAL 65

LOCOMOTION
Irene Markelic and Keyan Zahedi

xi
xii

AUTONOMOUS CLIMBING MOTIONS FOR CONNECTED 73

CRAWLER ROBOTS
Sho Yokota, Yasuhiro Ohyama, Hiroshi Hashimoto, Jin-Hua She,
Kuniaki Kawabata, Hisato Koabayashi and Pierre Blazevic

CONTROLLING AN ACTIVELY ARTICULATED SUSPENSION 81

VEHICLE FOR MOBILITY IN ROUGH TERRAIN
Siddhurth Sanan, Sartaj Singh and Krishna K Madhava

DESIGN AND CONSTRUCTION OF A ROPE CLIMBING ROBOT 91

Juan Pablo Martinez Esponda

DESIGN OF A NEW LEG MECHANISM FOR A WHEELED WALL 99

CLIMBING ROBOT
Yili Fu and Hejin Yang

DEVELOPMENT OF A CLIMBING ROBOT FOR INSPECTION OF 105

LONG WELD LINES
Jianzhong Shang, Bryan Bridge, Tariq Sattar, Shymal Mondal and
A h a Brenner

DEVELOPMENT OF A SEALING SYSTEM FOR A CLIMBING 115

ROBOT WITH NEGATIVE PRESSURE ADHESION
Carsten Hillenbrand, Daniel Schmidt, Karsten Berns, Tim Leichner,
Tobias Gastauer and Bemd Sauer

DEVELOPMENT OF A SUCTION TYPE MINIATURE CLIMBING 125

ROBOT WITH MINIMAL ACTUATORS
Muthu Veeruppan Vignesh and L. Karunamoorthy

DEVELOPMENT OF AN AMPHIBIOUS HEXAPOD ROBOT BASED 135

ON A WATER STRIDER
Soh Fujii and Taro Nakamura

DEVELOPMENT OF AN OM-DIRECTIONAL MOBILE ROBOT 144

BASED ON SNAIL LOCOMOTION
Kuniaki Satoh and Taro Nakamura

GAIT PARAMETER ADAPTATION TO ENVIRONMENTAL 153

PERTURBATIONS IN QUADRUPEDAL ROBOTS
Elena Garcia, Joaquin Estremera, Pablo Gonzalez de Santos and
M. Armada
 xiii

INTELLIGENT SPIDER WALKING ROBOT FOR ROUGH TERRAIN 161

Michael McCready, Liqiong Tang and Guwinder Singh Virk

KINEMATICS, SENSORS AND CONTROL OF THE FULLY 169

AUTOMATED FACADE CLEANING ROBOT SIRIUSC FOR
THE FRAUNHOFER HEADQUARTERS BUILDUNG, MUNICH
Norbert Elkmunn, Mario Lucke, Tino Kriiger, Dietmar Kunst and
Thomas Sturze

ON FOUR LEGS TOWARDS FLEXIBLE AND FAST LOCOMOTION 177

Cem Kara, Christian Heckhofi Thorsten Brandt and Dieter Schramm

ON THE DESIGN OF A FOUR-BAR MECHANISM FOR 185

OBSTACLES CLIMBING WHEELS
Antonio Gonzalez, Erika Ottaviano and Marco Ceccarelli

PATH PLANNING FOR THE “3DCLIMBER’ 193

Mahmoud Tavakoli, Lino Marques and Ani3al T. de Almeida

ROBOT FOR MOTION IN TUBE 203

Jatsun Sergey, Mishenko Vladimir and Jatsun Andrey

ROBOTRAIN AS SNAKELIKE ROBOTIC SYSTEM WITH 211

MINIMAL NUMBER OF DOF
V. E. Pavlovsky, V. V. Pavlovskyjr., N. V. Petrovskaya and
V. V. Evgrafov

SERVICING SOLAR POWER PLANTS WITH WALLWALKER 219

Ridha Aziz

STABILITY AND GAIT OPTIMIZATION OF A HYBlUD 226

LEGGED-WHEELED ROVER
Byron F. Johns and Ayanna M. Howard

TERRAIN-ADAPTIVE LOCOMOTION OF A WHEEL-LEGGED 234

SERVICE ROBOT USING ACTUATOR-BASED FORCE
MEASUREMENTS
Petri Virekoskiand Ilkka Leppanen

CONTROL OF QUADRUPED WALKING ROBOT BASED ON 242

BIOLOGICALLY INSPIRED APPROACH
Tae Hun Kang, Ig Mo Koo, Young Kuk Song, GiaLoc Vo,
Tran Duc Trong, Chung Min Lee and Hyouk Ryeol Choi
xiv

THE IMPROVEMENT OF STRUCTURAL AND REAL TIME 252

CONTROL PERFORMANCES FOR MERO MODULAR
WALKING ROBOTS
Ion Ion, Luige Vladareanu,Radu Munteanu jr. and Mihai Munteanu

Advances in Walking Robots

A BASIC VARIABLES SET BASED SCHEME OF ONLINE MOTION 263

PLANNING FOR HUMANOID ROBOTS
Jian Wang, Tao Sheng and Hongxu Ma

A HOPPING MOBLITY CONCEPT FOR A ROUGH TERRAIN 27 1

SEARCH AND RESCUE ROBOT
Samuel Kesner, Jean-Sdbastien Plante, Steven Dubowsky and
Penelope Boston

A PROPOSAL FOR BIPEDAL LOCOMOTION USING GYROSCOPIC 28 1

EFFECT
Pulkit Kapur, Rahul Mukhi and Vinayak

A SELF-ADJUSTING UNIVERSAL JOINT CONTROLLER FOR 289

STANDING AND WALKING LEGS
Axel Schneider, Bjom Fischer, Holk Cruse and Josef Schmitz

A STEP TOWARDS PNEUMATICALLY ACTUATED BIPED 299

LOCOMOTION: A BIO INSPIRED PLATFORM FOR STIFFNESS
CONTROL
Giovanni Muscat0 and Giacomo Spampinato

AUTONOMOUS BIPEDAL GAIT ADJUSTMENT UNDER 309

PERTURBATIONS
Lin Yang, Chee-Meng Chew and Aun-Neow Po0

DESIGN AND PROBLEMS OF A NEW LEG-WHEEL WALKING 319

ROBOT
Cristina Tavolieri, Erica Ottaviano and Marc0 Ceccarelli

STIFFNESS AND DUTY FACTOR MODELS FOR THE DESIGN OF 329

RUNNING BIPEDS
Muhammad E. Abdallah and Kenneth J. Waldron

CONSTRAINT BASED TRAJECTORY SIMPLIFICATION OF 340

FULL BODY TRAJECTORIES FOR A WALKING ROBOT
Hanns W. Tappeiner and Alfred A. Riui
 xv

FOOT PLANNING MOTION OF HUMANOID ROBOT RH-1 USING 347

LAG ALGORITHM
Mario Arbulu, Luis Cubus, Dmitry Kuynov, Pave1 Sturoverov and
Carlos Balaguer

GAIN PROPERTY FOR BIPED WALKING VIA LEG LENGTH 357

VAlU ATION
Tetsuya Kinugasu, Shoichi Miwu, Yunnick Aoustin and
Christine Chevullereuu

LEG CONTROL FOR CHANGING LOCOMOTION BETWEEN 365

LEG-TYPE AND WHEEL-TYPE BASED ON EFFECTIVE USE
OF TOTAL POWER
Tokuji Okadu, Wagner Tunuka Botelho and Toshimi Shimizu

MOVEMENT SIMULATION FOR MERO MODULAR WALKING 373

ROBOT
Ion Ion, Ion Simionescu, Adrian Curuj and Alexandru Murin

TRAJECTORY GENERATOR FOR RHYTHMIC MOTION CONTROL 383

OF ROBOT USING NEURAL OSCILLATORS
Weiwei Huung, Chee-Meng Chew, Geok-Soon Hong and
Nithyu Gmmssegurune

OBSERVER-BASED CONTROL OF A WALKING BIPED ROBOT: 393

STABILITY ANALYSIS
VincentLebusturd, YunnickAoustin and Frunck Plestun

OPTlMIZATION OF HUMANOID ROBOT MOTION DURING 402

ELEVATION OF AN OBJECT
Humed Ajubi Nueini and Mostufu Rostumi

POSTURAL STABILITY CONTROL FOR ROBOT-HUMAN 408

COOPERATION FOR SIT-TO-STAND ASSISTANCE
Viviune Pusqui, Ludovic Suintbuuzel and Philippe Biduud

RESEARCH ON UNDERACTUATED DYNAMICAL WALKING OF 417

3D BIPED ROBOT
Sheng Tuo and Mu Hongxu

ROTOPOD: A NOVEL APPROACH TO EFFICIENT LEGGED 427

LOCOMOTION
Dumiun M. Lyons
xvi

THE DESIGN OF A HUMANOIDAL BIPED FOR THE RESEARCH 435

ON THE GAIT PATTERN GENERATORS
Przemyslaw Kryczku and Chee-Meng Chew

THINKING ABOUT BOUNDING AND GALLOPING USING 445

SIMPLE MODELS
Kenneth J. Waldron, Jouquin Estremera, Paul J. Csonka and
Surya P. N. Singh

USING OPTIMIZATION TECHNIQUES FOR THE DESIGN AND 454

CONTROL OF FAST BIPEDS
Tobias Luksch, Karsten Bems, Katja Mombaur and
Gerrit Schultz

USING VIRTUAL MODEL CONTROL AND GENETIC ALGORITHM 466

TO OBTAIN STABLE BIPEDAL WALKING GAIT THROUGH
OPTIMIZING THE ANKLE TORQUE
Van-Hum Dau, Chee-Meng Chew and Aun-Neow Po0

WALKER SYSTEM WITH ASSISTANCE DEVICE FOR 475

STANDING-UP
Duisuke Chugo, Wataru Matsuoku and Kunikutsu Takuse

Advances in Humanoid Soccer Robots

A DISTRIBUTED EMBEDDED CONTROL ARCHITECTURE FOR 487

HUMANOID SOCCER ROBOTS
Carlos Antonio Calderon, Chungjiu Zhou, Pik Kong Yue,
Mike Wong and Mohan Rajesh Elara

DESIGN OF A HUMANOID SOCCER ROBOT: WUKONG 497

Qing Tang, Rong Xiong, J i m Chu and Xinfeng Du

FORMULATION OF DESIRED ZERO MOMENT POINT TRJECTORY 506

USING STATISTICAL METHOD
Lingyun Hu, Changjiu Zhou, Bi Wu and Tianwu Yang

LOCOMOTION CONTROL SCHEME FOR FAST WALKING 5 15

HUMANOID SOCCER ROBOT
Weerayut Sawasdee, Pusan Kulvanit and Thavidu Maneewam

OPTIMUM PERFORMANCE OF THE FAST WALKING HUMANOID 523

SOCCER ROBOT: EXPERIMENTAL STUDY
Pasan Kulvanit, Bantoon Srisuwan and Djitt Laowattana
 xvii

Supporting Technologies

A BIOLOGICALLY INSPIRED ARCHITECTURE FOR CONTROL 533

OF GRASPING MOVEMENTS OF AN ANTHROPOMORPHIC
GRIPPER
Sergio Varona Moya, Javier Molina Vilaplana,
Alejandro Linares Barranco, Jorge Juan Feliu Battle
and Juan Ldpez Coronado

A CONCURRENT PLANNING ALGORITHM FOR DUAL-ARM 541

SYSTEMS
Jen-Hui Chuang, Ting-Wei Chan and Chien-Chou Lin

A MODULAR APPROACH FOR CONTROLLING MOBILE ROBOTS 547

Kristian Regenstein, Thilo Kerscher, Clemens Birkenhofer,
Tamim Asfour, J. Marius Ziillner and Riidiger Dillmann

A SELF ORGANIZING NETWORK MODEL FOR CLAWAR SYSTEM 555

COMMUNICATION COEVOLUTION
Fabio P. Bonsignorio

AN APPROACH TO GLOBAL LOCALIZATION PROBLEM USING 565

MEAN SHIFT ALGORITHM
Giovanni Muscat0 and Salvatore Sessa

ASYNCHRONOUS LOCAL POSITIONING SYSTEM BASED ON 575

ULTRASONIC ACTIVE BEACONS AND FEED FORWARD
NEURAL NETWORKS
Pablo Estkvez, Juan Hernandez, Jose' Cappelletto and
Juan Carlos Grieco

CONTACT PROCESSING IN THE SIMULATION OF CLAWAR 583

Tamas Juhasz, Mykhaylo Konyev, Vadym Rusin and
Ulrich Schmucker

CREATING A GESTURE RECOGNITION SYSTEM BASED ON 591

SHIRT SHAPES
Pave1 Staroverov, Silvia Marcos, Dmitry Kaynov, Mario Arbulu,
Luis Cabas and Carlos Balaguer

DESIGN AND DEVELOPMENT OF MICRO-GRIPPING DEVICES 599

FOR MANIPULATION OF MICRO-PARTS
Z. W. Zhong, S. K. Nah and S. H. Tan
xviii

DESIGNING OF A COMMAND SHAPER USING MULTI-OBJECTIVE 607

PARTICLE SWARM ALGORITHM FOR VIBRATION CONTROL OF
A SINGLE-LINK FLEXIBLE MANIPULATOR SYSTEM
M. S. Alum. M. 0. Tokhi and M. A, Hossain

DETECTING SOUND SOURCES WITH THE HUMANOID ROBOT 615

RH-1
Pave1 Staroverov, Ricardo Martinez, Dmitry Kaynov, Mario Arbulu,
Luis Cabas and Carlos Balaguer

IN SEARCH OF PFUNCIPLES OF ODOUR SOURCE LOCALISATION 623

E. E. Kadur, G. S. Virk and C. Lytiridis

GA TUNED CLOSED-LOOP CONTROL OF SPRING BRAKE 632

ORTHOSIS
M. Saiful Huq, R a s h Massoud, M. Shaiful Alum and M. 0. Tokhi

HIDDEN MARKOV MODEL BASED FUZZY CONTROLLER FOR 642

FLEXIBLE-LINK MANIPULATOR
M. N. H. Siddique, M. A. Hossain, M. S. Alum and M. 0. Tokhi

HWSIL BY DEVELOPMENT OF SIX-LEGGED ROBOT SLAIR2 652

Sergiy Dzhantimirov, Frank Palis, UlrichSchmucker, Andriy Telesh
and Yuriy Zavgorodniy

IMPROVING POWER TO WEIGHT RATIO OF PNEUMATICALLY 662

POWERED LEGGED ROBOTS
Graham Mchtchey and John Billingsley

INTEGRATED INTELLIGENT MECHROBOT SYSTEM 670

Liqiong Tang and Guwinder Singh Virk

MCA2 - AN EXTENSIBLE MODULAR FRAMEWORK FOR ROBOT 680

CONTROL APPLICATIONS
Klaus Uhl and Marc0 Ziegenmeyer

MOTION ESTIMATION AND SELF-LOCALIZATIONBASED ON 690

COMPUTER VISION AND ARTIFICIAL MARKER DEPOSITION
Savan Chhaniyara, Kaspar Althoefer and Lakmuld Seneviratne

NEW STANDARDS FOR NEW ROBOTS 698

Gurvinder Singh Virk
 xix

PAFULLEL PARTICLE SWARM OPTIMIZATION FOR NETWORKED 708

CLAWAR SYSTEM COOPERATION
Fubio P. Bonsignorio

PERFORMANCEMETRICES FOR IMPROVING HUMAN-ROBOT 716

INTERACTION
Yiunnis Gutsoulis and Guwinder Singh Virk

REAL,-TIME COMPUTATIONALCOMPLEXITY OF THE 726

ALGORITHMSFOR A SINGLE LINK MANIPULATOR SYSTEM
M. A. Hossuin, M. N. H. Siddique, M. 0. Tokhi and M. S. Alum

SOFIUTARE AND COMMUNICATIONINFRASTRUCTUREDESIGN 736

OF THE HUMANOID ROBOT RH-1
Dmitry Kaynov, Mario Arbulu, Pave1 Sturoverov, Luis Cubus and
Carlos Buluguer

SPARBOT - A ROBOTIC FOCUS MITT TRAINING PLATFORM 744

Richard Stokes, Liqiong Tang and Ibruhim A. Al-Buhadly

ACTUATORS AND ORTHOSES TO ASSIST FES-CYCLING 752

Rusha Mussoud, M. Osman Tokhi, M. Shu.ul Alum and M. S. Huq

A NOVEL MINIATURE ATI'ITUDE MEASUREMENTSYSTEM FOR 761

CLIMBING AND WALKING ROBOTS
Guunglong Wung, Chunxi Zhang, Zhaoying Zhou and Rong Zhu
This page intentionally left blank
Plenary Introduction
This page intentionally left blank
 BIPEDAL HUMANOID ROBOTICS AND ITS APPLICATIONS

ATSUO TAKANISHI~
Department of Modern Mechanical Engineering
Waseda University, Japan

1. Introduction

Even though the market size is still small at this moment, applied fields of
robots are gradually spreading from the manufacturing industry to the others in
recent years. One can now easily expect that applications of robots will expand
into the first and the third industrial fields as one of the important components to
support our society in the 21st century. There also raises strong anticipations in
Japan that robots for the personal use will coexist with humans and provide
supports such as the assistance for the housework, care of the aged and the
physically handicapped, since Japan is one the fastest aging societies in the
world. Consequently, humanoid robots andor animaloid robots have been
treated as subjects of robotics researches in Japan such as a research tool for
humadanimal science, an entertainmendmental-commit robot or an
assistandagent for humans in the human living environment.

Over the last couple of years, some manufactures including famous global
companies started to develop prototypes or even to sell mass production robots
for the purposes mentioned above, such as SONY, TMSUK, ZMP, TOYOTA,
HONDA, etc. Most of those robots have two legs for its mobility . On the other
hand, Waseda University, where we belong to, has been one of the leading
research sites on bipedal walking robot and humanoid robot research since the
late Prof. Ichiro Kato and his colleagues started the WABOT (WAseda roBOT)
Projects and developed the historical humanoid robots that are WABOT-1 and
WABOT-2 done in the early 70s and 80s respectively.

One of the most important aspects of our research philosophy is as follows: By

constructing anthropomorphichumanoid robots that fimction and behave like a

3
4

human, we are attempting to develop a design method of a humanoid robot

having human two legs to coexist with humans naturally and symbiotically, as
well as to scientifically build not only the physical model o f a human but also
the mental model of it .From the engineering view point. Based upon the
philosophy, I and my colleagues have been doing researches on bipedal
humanoid robots. In my plenary speech I will introduce the research philosophy
of bipedal humanoid robotics introducing current bipedal walking robots
WABIAN-2R and WL-16RIII as shown in Fig. 1 and 2 respectively, the design
and the control of the robots and its applications collaborating with robotics
companies

Fig. 1 Bipedal Humanoid Robot Fig. 2 Bipedal Walking Vehicle WL-

WABIAN-2R that walks with 0.96 s 16RIII that carries 75 kg human and
step time and 0.5 m step length. walks up/down on 0.25 height stairs.
 CLIMBING UP THE WALL

JOHN BILLINGSLEY
University of Southern Queensland
Toowoomba, Australia

Reminiscences cover over two decades of legged robot research, From Robug I, a six-
legged walker that was overweight, the story leads through a succession of wall-climbers
both simple and elaborate, some of which have found commercial applications in the
nuclear industry.

1. Introduction
The robotics research lab of Portsmouth Polytechnic, as it was then, was
transformed in June 1985. Of course, work on the Craftsman Robot and
manufacturing cells continued, but in that month Arthur Collie joined the Group
with a Royal Society grant to indulge his passion for legged robotics. The first
project resulted in Robug I, an overweight six-legged machine with a body in the
shape of a coffin. Later robots such as Robug I1 were designed with weight very
much in mind, although they still favoured pneumatic actuation.
While some projects pursued intelligence (or rather cunning), others such as
Zig-Zag sought to improve mechanical simplicity. However the nature of the
research was transformed when the nuclear industry took an interest.
Early collaboration with Arthur had been through his role as Development
Manager of Turnright Controls Ltd, part of the Tube Investments Group. The
interest then had been in controls for domestic cookers, a far cry from walking
robots. A realignment of the Tube Investments structure led to a management
buy-out and Turnright’s transformation into a new company, Portsmouth
Technology Consultants, or PorTech for short.
The new company needed products to sell. When the nuclear industry
expressed a readiness to place an order for a robot to enter a reactor, the
opportunity was seized and a succession of Neros resulted.
By 1992 I had had my fill of winter arthritis and all-too-brief English
summers. I deserted Portsmouth for the warmth of Queensland. At the
University of Southern Queensland on top of the Great Dividing Range the work
on legged robots continued.

5
6

Within a year, Michael Bellmann had completed Hydra, a 'somersaulting'

wall-climber. Michael Rook developed the Toad for his master's degree and by
1997 Sam Cubero had completed STIC, a lightweight four-legged ceiling-
walker.
Some projects are waiting in limbo, half completed. Nantha Kumar's roller-
skating robot still occupies considerable space in the lab, but is a striking icon
that gets photographed on Open Days.
Graham McLatchey has pulled most of the legs off Robug 4, which we
scavenged from the demise of Portech a few years ago. He is seeking to improve
the efficiency of the actuation, but will present his own account of it.

2. The first Robug

We did not enter the task blindly.
Arthur's funding was by the Royal Society under their industrial fellowship
scheme, an innovative initiative to increase the exposure of universities to
industry. The fellowship gave the full facilities of the Royal Society for its two-
year duration, paying his salary at a reduced amount. Funding for the project
itself came from SERC. Arthur relates that the SERC representative said, very
grudgingly, "I suppose under the circumstances we shall have to give you a
grant."

Figure 1 . Robug 1 was shaped like a coffin, filled with circuit boards
There was also a small grant from the Ministry of Defence on the grounds
that the machine could be used either as a sort of mine clearing machine or as a
single mission device. As a mobile antitank mine for use in rugged terrain it
would lie in wait in a ditch, clear of the mine clearing flails, then jump up, run
towards the sound of vehicles and go out in a suicidal blaze of glory.
 7

Accordingly we set off to inspect the DARPA legged robot in Columbus,

Ohio. Some massive engineering brute force made the task look easy. It was
not.
Robug 1's body was shaped rather like a coffin. Inside it, six single-board
BBC Micros (the embedded processor of choice at that time) provided the
computing power, one for each leg. It is interesting to compare the sizes of trial
legs of Robug and DARPA. Their budgets were equally disproportionate.

Figure 2. Legs of different sizes, DARPA and Robug I

Since it must be self-powered, it was intended that pneumatic power should
eventually be provided by an explosive gas generator and an accumulator.
Pneumatic actuation would also give the compliance that would allow the load to
be shared evenly between the legs.
Much was learned from Robug I. Compliance and stability are uneasy
bedfellows. Without an inertial sensor, Robug's early steps were somewhat
faltering. We also learned how easy it is to underestimate the load each leg
would have to bear, not only from movement of the centre of gravity but also
from inertial forces.
In those days, before mechtronics became recognised as an integrated
discipline, we discovered a fault of human nature. During some holidays, the
project's programmer had been left in charge of the machine. We returned, to be
told that the time had been spent writing a wonderful new algorithm that
identified dynamic errors in one of the angle sensors and adaptively
compensating for them. Arthur showed great restraint in the way he told the
8

young lady in question that using an Allen key to tighten the sensor's grubscrew
would have been an altogether better solution.

3. RobugII
In 1988, another grant allowed the work to continue. Armed with plenty of
hindsight, we aimed for low weight. To improve stability, the robot followed an
insect-based scheme with its knees higher than its centre of gravity. Wall-
climbing looks more dramatic than walking, but it allows stability to be
simplified even further. A "Shufflebottom" version with two am/legs with
pneumatic grippers, mounted on a base with two more suckers, showed that the
principle worked. Anothertwo legs were added to make the full Robug II.

Figure 3. Shufflebottom and Rob

With 'belly suckers' augmenting the legs, a strange but safe gait was
possible. Each foot was released and advanced in turn, then the body was
released and heaved forwards. Strategies of exploring for a safe foothold
impressed the viewers of BBC'elev Vision's 'Tomorrow's World'.

Figure 4: Robug II on "Tomorrow's World" and climbing on a building

 9

We then wanted to make a transition from floor to wall, or from wall to

ceiling, so the body gained a hinge so that the front part could rear up. Bing
Lam Luk joined the project and took the software in a new direction.

4. Zig-Zag
A single pneumatic cylinder flexes a square jointed frame to become a rhombus.
Four suckers at the corners hold it to a wall. The grippers on one side are
released, the shape is flexed and Zig-zag has taken a step. Arthur boasted this as
the simplest form of walking robot. However it has the disadvantage that the
unattached corners merely 'flop' against the wall and it cannot operate on an
overhang. Nevertheless in 1990 Arthur and Bing were proud to take it to the
first Robot Olympics, where it won its event on a high vertical wall.

F i y r e 5 . Zig-zag with Arthur and Bing at the 1990 Robot Olympics

5. Hydra
By now I had made my escape to Australia. I had an idea for a simple walker
that could make the transition from floor to wall. It had just two large suction
feet and one joint. It would stand doubled over, like an athlete gripping his
10

ankles, then straighten into a series of 'flicflac' somersaults that would take it
across the floor. The addition of swivelling joints in the 'ankles' would give
navigation and with careEul positioning the Hydra could transfer to a wall.

Figure 6 . The principle of Hyda

Michael Bellmann joined me in Toowoomba and within months Hydra was
completed as part of his project work for TU Muenchen.

6 . Toad
I wanted a more versatile competitor for Zig-Zag and the result was Toad. Two
front suckers are mounted at the ends of a horizontal bar. Two rear suckers are
mounted at the ends of a second bar. The two bars are joined by a link that can
be extended or shortened by a p n e ~ a t i cylinder.
c

Figure 7. Toad I, and a simpler version running on the ceiling on Tomorrow's World
If the front left foot is released and the link is extended, Toad takes one step
forwards. The foot grips again, the rear left foot is released and the li& retracts
to take the next step. This is repeated on the right-hand side and Toad has
advanced by a full pace.
 11

Some simulation showed that the rear pivot of the cylinder must be behind
the feet to ensure that the rear legs are stable in lining up with the front ones.
Extra actuation was added to twist the link, so that when a foot is released it can
be actively lifted or pressed against the surface to take a new grip.
Michael Rook developed the machine for his Master's project, and
Tomorrow's World viewers saw a Toad running across the studio ceiling.

7. Nero
Meanwhile in England, Portech had taken up the challenge of making a robot for
the nuclear industry. The first robot task was to climb on a reactor pressure
vessel in order to put a line of temperature sensors in place and Nero I had this
single purpose.
Whereas we had struggled to incorporate intelligence into Robug, the last
thing the clients wanted was a robot that moved on its own initiative. Each pace
must be verified by an operator. If a mishap occurred and the robot had to be
retrieved by a human, reactor shut-down costs would amount to six figures.
The design was in the form of two rectangular frames, each with a gripper at
each corner. One was long and thin, inside which a carriage could move up and
down. The second was square, also carried four grippers, was mounted on the
carriage and could be made to swivel. Frames were released and lifted in turn,
while the carriage operated either to carry the smaller frame forward, or the
larger frame if that was lifted.

Figure 8. Nero and its working environment

The success of Nero 1 led to several M h e r Neros. Soon they were working
upside-down in ducts of flowing hot gas, with grinding wheel attachments for
tidying up the metalwork.
12

8. STIC
Sam Cubero was determined to make a rival for Robug, to be lighter and more
agile. Once again the actuation was to be pneumatic.

Figure 9. Sam Cubero proudly demonstrated STIC

A vital part of the control was in the valves used to control the cylinders.
Poppet valves released huge amounts of air as they controlled pressure with
mark-space operation. Could a proportional gas valve be devised that would
permit more sophisticated control? A combination of the Coanda effect and
planar solenoids seemed to be the answer. Indeed STIC walked on a ceiling,
rather more steadily than on a floor, and is now in retirement in Sydney's
Powerhouse Museum.

9. Robugs I11 and IV

The collaboration between Portech and Portsmouth University now gained
strength with two major projects from the European Community. To some
extent, the Robugs were seen as a demonstrator for a new communication
protocol. robugIV has two dozen processor, three for each of its eight legs.

Figure 10. Robug 4

A few years ago, the supply of orders for walking robots dried up. Although
Portech had diversified into the supply of general mechatronic test equipment, its
 13

finances were found to be no longer viable. The consolation was that we were
able to buy Robug IV for the University of Southern Queensland.
With eight legs, concerns for the robot's 'standing strength' should be
minimal. Indeed there is video of someone sitting on Robug 111. However
weight was not so strictly constrained and when operated at a laboratory pressure
of 4 bar, as opposed to the design pressure of 10 bar, actuator force again
becomes an issue. Graham McLatchey has been working on solving it.

10. Conclusion
As robots become more commonplace, public interest is waning. Autonomous
vacuum cleaners and lawnmowers have not taken the market by storm. When a
robot has legs, however, it still seems to keep its cachet. Walking and climbing
robots will continue to be great fun for many years.

Acknowledgment
Arthur Collie has given me substantial help in refreshing my memory of those
early days. Bing Luk has also retained an archive.

References
1. The development of a pneumatically powered walking robot base, A A
Collie, J Billingsley, L Hatley, Proc IMechE C377/86, Conf. UK Research
in Advanced Manufacture, London, Dec. 1986, pp137-143.
2. Design and performance of the Portsmouth climbing robot, A A Collie, J
Billingsley, S F Mo, E von Puttkamer, Proc 7th International Symposium on
automation and robotics in construction, June 1990, Bristol.
3. Design and performance of the Portsmouth climbing robot, A A Collie, J
Billingsley, S F Mo, E von Puttkamer, Mechatronic Systems Engineering, 1,
pp 125-130
4. A climbing robot with minimal structure, J Billingsley, A A Collie, B L Luk,
Proc IEE conference Control 91, March 1991, Edinburgh, 2, pp 8 13-815.
5. Robug 11: an intelligent wall climbing robot, B L Luk, A A Collie, J
Billingsley, IEEE conference on Robotics and Automation, April 1991,
Sacramento, USA, pp 2342-2349.
6. Real time software control system for the Nero wall-climbing robot, B L
Luk, A A Collie, J Billingsley, T White, N Bevan, Proc 4th Euromicro
Workshop on Real Time Systems, Athens, June 1992, pp 74-78
7. An improved Wall Climbing Robot with Minimal Actuation, J. Billingsley,
A A Collie, M J Rook, Proc 1st IFAC International Workshop on Intelligent
Autonomous Vehicles, Southampton UK, 18-21 April 1993, pp 14-19.
14

8. An Articulated Limb Climbing Vehicle with Autonomous Floor to Wall

Transfer Capability, B L Luk, A A Collie, N Bevan, J Billingsley, Proc 1st
IFAC International Workshop on Intelligent Autonomous Vehicles,
Southampton UK, 18-21 April 1993, pp 20-24.
9. Toad - a wall-climbing robot, J Billingsley, M J Rook, A A Collie,
Australian Robot Association Conference, Robots for Competitive
Industries, Brisbane, July 14-16 1993.
10. Design Aspects of a Leg for a Walking Robot, J Billingsley, A Collie, Proc
Mechatronics and Machine Vision in Practice, Toowoomba Australia,
September 13-15 1994, pp 162-165.
11. Automatic surface transition adaptation for a quadrupedal space frame
robot, S N Cubero, J Billingsley, Proc Second International Conference on
Mechatronics and Machine Vision in Practice, Hong Kong, September 12-
14 1 9 9 5 , 113-118.
~~
12. A novel proportional gas valve for mechatronics applications, S N Cubero, J
Billingsley, Proc Second International Conference on Mechatronics and
Machine Vision in Practice, Hong Kong, September 12-14 1995, pp 179-
184.
13. Command and compliance of a leg for a walking robot, J Billingsley, A A
Collie, Proc Second International Conference on Mechatronics and Machine
Vision in Practice, Hong Kong, September 12-14 1995, pp 256-259.
14. Automatic control of a surface adapting, four legged wall climbing robot, S
J Cubero, J Billingsley, Mechatronics 96 With Mechatronics & Machine
Vision in Practice 96, University of Minho, Geuimaraes Portugal,
September 1996 pp 1.135-1.142.
15. S Cubero, J Billingsley, Force, Compliance, and Position Control for a
Space Frame Manipulator, 4th Mechatronics and machine vision in
practice, Toowoomba; Australia, Sep 1997. pp. 124-130.
16. N K Kanesen, J Billingsley, A Roller Skating Robot, Mechatronics and
Machine Vision, Research Studies Press, England, September 2000, ISBN
0-86380-261-3, pp 283-292.
17. McLatchey, G.J. and Billingsley J, Force and Position Control Using
Pneumatic Cylinders, 9th International Conference on Climbing and
Walking Robots, 2006, Belgium.
 From Micro to Nan0 and Swarm Robots

Heinz Worn, Ramon Estaiia, Heiko Hamann, and Marc Szymanski

Institute for Process Control and Robotics, Universitat Karlsruhe,
Engler-Bunte- Ring 8, 76131 Karlsruhe, Germany
E-mail: {woern, estana, hamann, szymanski}@im.uka.de
wwwipr. ira. uka. de

Current research in Micro, Nan0 and Swarm Robots as results of the European
projects Miniman, MiCFloN and I-SWARM will be presented. First, the
design and the control of 5 t o 10cm3 sized mobile micro robots with five degrees
of freedom will be shown.They can handle miniaturized parts as for example an
optical component or a biological cell with a size in the micrometre-area with
an accuracy of lOOnm under a microscope or a raster-electron microscope. Sec-
ond, the design and the control of a km3-sized mobile untethered micro robot
will be demonstrated. Here, the robot consists of five parts: the Piezzo locomc-
tion module, the micro control unit, the communication unit, the navigation
system and the micro gripper. The mobile robot can be guided and positioned
in an arena with an accuracy of 5 micrometre and can be programmed and
controlled over the wireless communication unit. Third, the design and the
control of 3 x 3 x 3 mm3 sized micro-/nanorobots with 2 degrees of freedom
will be presented. The transmission of energy and the communication between
the robots is realized via infrared. The robot controller is fully integrated and
has limited functionalities. Via basic sensors communication functions and ele-
mentary rules and behaviours the micro robot can act in a swarm consisting of
hundreds and thousands of robots. Future applications could be monitoring-,
inspection-, exploring-tasks etc. of big areas or objects.

Keywords: Micro Robots, High precise Navigation, Swarm Intelligence, Micro

Handling

1. The Miniman Project

This section gives a brief overview over the EU-sponsored Miniman
project. The main result of the research activity consists of the design and
fabrication of a set of components which enhance and make feasible the
micro manipulation and end-effectors, force and tactile sensors, gripper sys-
tems and their integration in a flexible micro robot system, thus resulting
in an essential tool for very different micro manipulation tasks. The micro
manipulation workstation comprises a few versions of robots with different

15
16

overall size; the robot prototypes consist of a mobile positioning unit with
two translatory and one rotational degree of freedom (DOF), which car-
ries a manipulation unit with a micro tool. The robots can operate at high
speed (up to 30 mm/s) and have a motion resolution of about 10 nm. In
Fig. 1 two types of Miniman micro robots are depicted.

a
Fig. 1. Two types of the Miniman micro robot as a result of the project.

The positioning units of the robots consist of three tube-shaped piezo legs,
which can be actuated independently by a control software, which includes
all necessarry algorithms for moving the robot over the smooth glass plate
of the working plate (300 x 200 mrn in size). The movement is based onto
the well established sli~stick-principleused already in other micro robot
projects like Nanowa1ker.l The Miniman is controlled via an 6D space
mouse and can be used for example as an intelligent tweezer in a SEM (see
Fig. 2) as well as a tool for micro lens assembly and cell inspection.
As already mentioned, specially designed micro robots can also act as a flex-
ible assembly facility for prototype micro systems in a Scanning Elecctron
Microscope (SEM), as probing devices for in-situ tests in various appli-
cations or just as a helpful teleoperated tool for the SEM operator when
examining samples. Several flexible micro robots of this kind have been de-
veloped and tested. Driven by piezoactuators, these few cubic centimeters
small mobile robots perform manipulations with a precision of up to 20nm
and transport the gripped objects at speeds of up to 3cm/s.
As described in this section, several new tools have been developed to mas-
ter the requirements of the micro world, i.e. micro lens assembling. The
system is now in a state for continuous usage, and it is a helpful tool for
various micro manipulation tasks. The results obtained give clear indica-
tion about parameters and steps which should be important for further
miniaturisation, for example the use of miniature actuators. At the same
 17

Fig. 2. A Miniman Gripper (left) and a Miniman needle as a helping hand in a SEM.

time, the system proves that the development of mobile micro robots is a
promising approach to realise very small and flexible tools useful for differ-
ent applications. By means of its intuitive teleoperation mode, the system
enables the user to work in the micro world.

2. The MiCRoN Project

2.1. I n t ~ ~ u ~ ~ ~ o n
The EU-sponsored MiCRoN project aimed at the development of a com-
pletely new micro robot system based on flexible mobile, 1 em3 sized robots
acting autonomouslyand and fully untethered. Eight European project part-
ners started in March 2002 with the development of a system based on the
results from the Miniman Project. The MiCRoN project’s outcome is
a major contribution in the field of micro mechatronic components. The
new micro robot is using low voltage piezo actuators and hybrid on-board
electronics. Several micro tools have been developed:

~illimetre-sizedgrippers for the 3D assembly of meso-scale objects

e A robot-mounted micro syringe chip for the injection of substances
into living cells
e A robot-mounted micro syringe chip for the injection of substances
into living cells
e AFM tools for standard AFM imaging and using functionalised
nano tips for biological experiments.
18

The robot actuators have been developed using a novel rapid prototyping
process for multilayer piezoceramics; the robot hardware consists of an on-
board VLSI electronics module. Another major contribution of the project
are methods for wireless micro robot operation. A power floor has been
built as a working prototype, which transmits electrical energy to mobile
units operating on a 250 x 250 mm2 area. The developed infrared com-
munication schemes are used for controlling a small group of robots. An
advanced control system has been realized including an implementation of
the progressive Kalman Filtering Theorem. This software package includes
all software needed for robot control, navigation, planning, simulation and
user interfacing. For the integrated on-board vision system, a camera, which
can be mounted on mobile micro robots, has been developed into a working
prototype. This system can generate 3D data of micro manipulation sce-
narios. The computer vision system developed offers a broad range of stable
recognition algorithms for micro handling applications. In the field of object
localization, the global localisation system represents a major achievement.
This system has reached final prototype state with a position resolution
of about 5 microns over the complete size of the workspace. The system
components key figures can be summarized as follows:

0 Mobile micro robots: platform with holonomic three degrees of free-

dom (DOF) x, y and 4, equipped with wireless powering,micro grip-
pers, Syringechip for cell manipulations, on-board Atomic Force
Microscope (AFM)-tools, piezoactuated miniature motor, Onboard
electronics module and wireless IR-based communication module
0 Handling experiments: multi-robot scenarios for cell-handling ex-
periments, cell-injection experiments and involving the 3D assem-
bly of meso-scale products and objects using a micro soldering pro-
cess
0 Vision and Position Sensors: These are sensor systems which pro-
vide the global localisation, object recognition, scene interpretation
and local images to the system
0 System control: this is the brain of the complete system

Together with the high resolution navigation system,2 a highly reliable and
precise handling tool is available (see Fig. 3).
More informations about the MiCRoN and the Miniman robot can be
found at http://microrobotics.ira.uka.de/
 19

Fig. 3. The visionary project objectives for the MiCRoN project.

3. The I-SWARM Project

The EU-sponsored I-SWARM project aims to develop and build a robot
swarm of up to 1,000 Each one equipped with limited, pre-rational
on-board intelligence. These swarm should be capable of performing tasks
that are not possible with either a single micro robot, or with a small
group of micro robots. The expected outcome shall be an observable self-
organisation effect in the robot swarm similar to that seen within ecological
systems like ant states, bees colonies and other insect aggregations. Further
benefits are a greater flexibility and adaptability of the swarm to the en-
vironment, robustness to failures, etc. utilizing emergent effects. Within
the project the suitability of the strategies of animals and especially insect
colonies and swarm intelligence for managing micro robot swarms are in-
vestigated. One of the major goals of the project is to transform knowledge
of eusocial insect behaviour, into the swarm intelligence of robots and to
apply this to the completely new swarm of micro robots.

During the I-SWARM project, the Jasmine robot was designed and fab-
.
ricated for the software test purposes (see Fig. 4, www swarmrobot.org). It
has a size of 3 x 3 x 3 em3 and supports almost all concepts the I-SWARM
robot uses. Equipped with six infra-red emitters and receivers the robot is
a perfect platform due to its excellent communication abilities, which are
essential for swarm robot research. Different kinds of bio-inspired swarm
algorithms have been implemented and investigated with Jasmine robots.

The final I-SWARM micro robots with a size of 3 x 3 x 3 mm3 have

a design depicted in Fig 5 . The locomotion platform has three legs based
on a piezoelectric polymer actuator material. A vibrating needle utilizing
20

Fig. 4. Swarm of Jasmine robots.

the same actuator material acts as touch sensor. An ASIC comprises a 8051
based micro controller kernel, the power managment including amplifiers,
the communication interfaces for infra-red and the solar cells. The ASIC is
also the robot’s backbone and is connected via an FPC with the four channel
infra-red communication device and the solar cells. The solar cells deliver
the energy for the whole system and are also used as an ego-positioning
sensor.
This robot design is far beyond the current state-of-the-art micro robot
research. Indeed those micro robots may have a worse resolution regarding
the manoeuvrability and controllability than the current micro robots. But
there is no state-of-the-art micro robot swarm reaching the ~ - S ~ A R
constraints. Further more, the computational, communicational and senso-
rial constraints the on-board micro controller and sensors dictate lead to a
more basic robot swarm research. Swarm robotics will benefit from this fun-
dament research, as fancy algorithms using a whole personal computer to
control a single swarm robot could not be implemented on the ~ - S ~ A R
robots. The self-organisation abilities have to clearly rise from the pure
number of robots in the swarm and simple but intelligent algorithms.
Several bio-inspired algorithms like tropholaxis strategies have been trans-
ferred from nature to artificial swarms and tested with the J u s ~ ~ robot
ne
or in ~imulation.~ A motion description language (MDL2e) for rule based
swarm behaviour has been designed that is similar to a regular language.6
The description language is used to describe the robots behaviour in a sim-
 21

Fig. 5 . The design of the final 3 x 3 x 3 mm3 sized I-SWARM robot.

ulation and could also be directly transferred to the robot. The language
describes the robot controller based on primitive action so called atoms.
The description language is also used to generate controllers for emergent
behaviour using the paradigm of Genetic Progrummzng.

3.1. Modeling Swarm Robotic Systems to Support the

~ o n t ~ l l Design
er
Designing and implementing artificial self-organizing systems is in general
a challenging task since they typically behave non-intuitive and only lit-
tle theoretical foundations exist. Predicting a system of many components
with a huge amount of interactions is beyond human skills. A connected
complex of problems is the micro-macro link - determining the global be-
havior resulting from local behavior patterns and vice versa.? This applies
also to swarm robotic systems in particular. The currently common use of
simulations for design support is not satisfying, as it is time-consuming and
the results are most likely suboptimal.
The problem is to find an effective and efficient control algorithm for the
robot swarm to accomplish a predefined task. To minimize the complexity
of the entire system and due to the limited capabilities of an individual
swarm robot, the development targets simple rules and, in an allusion to
nature, one hopes for emergent behavior of the robot group that leads to
the solution of the given task. However, the design of self-organizing and
emergent behavior is very difficult. There are no general methods available
that support the engineer in designing and optimizing the control algorithm.
First steps to establish such a theory of self-organizing behavior in swarm
22

robotics have been proposed The fundamental substance of

these approaches is a model that approximately predicts the behavior of
the swarm. However, the prediction of emergence is considered to be hard
or even impossible as a contradiction in terms.11J2
In ongoing work the authors expand the set of available models by an ana-
lytical approach that describes space continously and explicitly rather than
approximating it discretely. In many self-organized swarms the effec-
13114

tivity of the behavior depends primarily on spatial inhomogeneities that

are hard to approximate by a discrete approach. Thus, it can be inferred
that an explicit and exact representation of space is useful in these cases.
Our model describes, first of all, the probability density function of the
robot's positions. This macroscopic description is derived from a micro-
scopic description of an individual robot's movement. This is done by re-
garding the swarm robotics system as a system showing Brownian mo-
tion with drift and using statistical physics to derive a model. Hence, the
micro-macro link is solved almost for free by established physical methods.
However, currently our model depends on heuristic reasonings as well to
model the communication by means of stigmergy or direct communication
because physics does not provide the theory for autonomous and explicitly
communicating particles.
At first, we model microscopically the movements of an individual robot
using a Langevin equation of the form
r = A(r) + C(r)F(t), (1)
with a noise term F(t), A(r) describing the deterministic proportion of
the robot's motion, and C(r) the nondeterministic proportion. From this
equation, a partial differential equation (PDE), the Fokker-Planck equation,
can be derived modeling the swarm robotic system macroscopically.
Omitting the d e r i v a t i ~ n l ~here,
y ~ ~the Fokker-Planck equation under certain
assumptions is

with the product f ( r ,t)dr,dr, determining the probability of encountering

a robot at position r within the rectangle defined by dr, and dry at time t.
See Fig. 6 as an example of f in an aggregation scenario.
We model direct or indirect communication between the robots, e.g. virtual
pheromones, by a potential field on which A(r) and C(r) depend. The de-
velopement over time of this potential field can be a PDE as well extending
the model to a system of PDEs.
 23

Once the model is derived it is possible to predict the swarm behavior

for a given set of parameters. If the algorithm is properly parameterized
it is feasible to scan whole parameter intervals to find regions of optimal
behavior. For a detailed description and a positive validation of two special
scenarios against a simulated robot swarm see13$14

Fig. 6. The robot density over space for an aggregation scenario as an example.

4. Summary and Outlook

In this paper we described the development of three different kinds of micro
robots. With the 5 - 10crn3 Miniman micro robot, it was proved that the
development of small micro manipulation units is a positive approach to
realise tools that can handle parts of the micro world. The lm3sized un-
tethered MiCRoN robot shows the higest miniaturisation of a micro robot
equipped with tools at the moment It enables autonomous cell handling as
well as a new kind of micro part assembly. The results of the I-SWARM
project are expected to have a huge impact in the longer term: realisations
of robotic swarms are conceivable in diagnostic tasks in a medical or techni-
cal context. Such a swarm could be inserted into an engine, pipeline system
or patient, gather distributed data, diagnose faults or illnesses, transmit
information to the outside and act accordingly. However, methods investi-
gated in the project can also be applied in the near future: self-organisation
and configuration becomes increasingly interesting for technical devices or
systems.
24

References
1. S . Martel, M. Sherwood, C. Helm, W. G. de Quevedo, T. Fofonoff, R. Dyer,
J. Bevilacqua, J. Kaufman, 0. Roushdy, and I. Hunter, Three-legged wireless
miniature robots for mass-scale operations at the sub-atomic scale, in roc. of
the ZOO1 IEEE Int. Conf. on Robotics and Automation, ed. E. SahinLNCS
3342 (Springer Verlag, Berlin, Germany, 2001).
2. R. Estafia and H. Worn, Moire-based positioning system for micro robots,
in SPIE’s Int. Conference on Optical measurement Systems for Industrial
Inspection 111,(SPIE, 2003).
3. R. Estafia, M. Szymanski, N. Bender and J. Seyfried, Towards a Real Micro
Robotic Swarm, in Ant Colony Optimization and Swarm Intelligence. 4th
International Workshop, ANTS 2004, ed. M. D. et al.LNCS 3172 (Springer
Verlag, Berlin, Germany, 2004).
4. J. Seyfried, M. Szymanski, N. Bender, R. Estana, M. Thiel and H. Worn, The
I-SWARM project: Intelligent Small World Autonomous Robots for Micrc-
Manipulation, in Swarm Robotics, ed. E. SahinLNCS 3342 (Springer Verlag,
Berlin, Germany, 2005).
5. T. Schmickl, C. Moslinger and K. Crailsheim, Collective perception in a
robot swarm, in Proceedings of the Second Swarm Robotics Workshop, eds.
E. Sahin, W. M. Spears and A. F. T. Winfield, LNCS, Vol. 4433, 2007.
6. M. Szymanski and H. Worn, JaMOS - A M D L ~ based E Operating System for
Swarm Micro Robotics, in IEEE Swarm Intelligence Symposium, eds. Y . Shi
and M. D. (Ed.), 2007.
7. D. Yamins, Towards a theory of ”local to global” in distributed multi-agent
systems, in AAMS’05, Utrecht, Netherlands, 2005.
8. A. Martinoli, K. Easton and W. Agassounon, Int. Journal of Robotics Re-
search 23, 415 (2004).
9. S. Kornienko, 0. Kornienko and P. Levi, Swarm embodiment - a new way
for deriving emergent behavior in artificial swarms, in Autonome Mobile Sys-
teme, eds. P. Levi, M. Schanz, R. Lafrenz and V. Avrutin, 2005.
10. K. Lerman, C. Jones, A. Galstyan and M. Mataric, Int. J. of Robotics Re-
search 25, 225 (2006).
11. V. Darley, Emergent phenomena and complexity, in Artificial Life 4 , eds.
R. Brooks and P. Maes, 1994.
12. C. Miiller-Schloer, Organic computing - on the feasibility of controlled emer-
gence, in CODES+ISSS’04, Stockholm, Sweden, 2004.
13. H. Hamann and H. Worn, A space- and time-continuous model of self-
organizing robot swarms for design support, in First IEEE International Con-
ference on Self-Adaptive and Self-organizing Systems (SASO ’07)’ Boston,
USA, July 2007.
14. H. Hamann and H. Worn, An analytical and spatial model of foraging in a
swarm of robots, in 2nd SAB 2006 International Workshop, eds. E. Sahin,
W. Spears and A. Winfield, LNCS, Vol. 4433 (Springer, April 2007).
15. H. Haken, Synergetics - an introduction (Springer, 1977).
16. N. G. van Kampen, Stochastic processes in physics and chemistry (North-
Holland, Amsterdam, 1981).
 CLIMBING ROBOTS FOR NONDESTRUCTIVE TESTING:
HISTORICAL PERSPECTIVE AND FUTURE TRENDS

BRYAN BRIDGE’
Go-Director, Research Centre for Automated and Robotic NDT
Faculty of Engineering, Science and the Built Environment
London South Bank University, United Kingdom

1. Introduction

It is perhaps not well known that mobile robots were already deployed in
Nondestructive Testing (NDT) in the early 1970’s, well before fixed robots
came into use on manufacturing plant. The need for NDT robots first arose with
the growth in exploitation of the world’s oil and gas supplies. Vast length of
pipeline are involved in transporting these supplies from source to multiple
destinations. Sediment and chemicals in the unrefined products cause rapid
cause inner wall thinning from both erosion and chemical attack. Pipe rupture at
just one point, anywhere along a pipe can cause a major environmental disaster
so there was always a need to inspect complete lengths of long runs of pipe. The
handling of inspection sensors by human operators would thus always involve
vast and costly numbers of personnel. This situation begged for sensor handling
by mobile robot. A second reason for robotic deployment arose from the fact
that many pipelines were buried underground, under concrete or ran along the
sea bed before rising to the surface (risers). Human placement of sensors is
impossible in these cases.

2. Pipelines

Five kinds of mobile and climbing robots were developed to deal with the above
pipeline inspection problems. By definition a generic requirement will be that
the robot will be adaptable in some way to curved surfaces.

25
26

(i) Pipeline pigs were cylindrical vehicles tailored to fit snugly inside pipes
and move along a pipe under pressure hand having a collar of electromagnetic
or ultrasonic sensors which automatically scanned the internal pipe walls for
corrosion and erosion thinning. British Gas invested several Million pounds (at
todays’s prices) on the R &D of such a vehicle. It won a major prize for
industrial innovation and was exhibited at the Science museum, which was
testament to the perceived importance of and need for this technology.

(ii) Internal pipe crawlers were wheeled vehicles designed for pipe weld
inspection. They carried panoramic X ray tubes inside pipes with co-moving
detectors (film, phosphor screens image intensifiers etc) on the outside. This
arrangement, only possible when external access to the pipe was possible
allowed the preferred single wall through transmission inspection method to be
deployed. The first crawler to be made small enough to fit inside 8 inch pipes,
produced by Oilfield Inspection Services at Great Yarmouth, was reported on
the front page of one edition of the Financial Times in the late 1970’s.

(iii) External pipe crawlers. These robots, usable when there is external access
to the pipe surface, were developed in the early eighties. Usually they had
limited movement in which ultrasonic sensors performed orbits round
circumferential weld. At is simplest the motion was simply achieved by a chain
drive and the robot moved manually from one weld to the next. The pioneer of
this technology in Europe was the Danish Welding Institute (Densk Norsk
Veritas).

(iv) Small wheeled vehicles which could move both around and along pipes
with magnetic adhesion emerged soon after (iii) as an alternative way of
deploying ultrasonic sensors on accessible pipe surfaces.

(v) Submarine (swimming) inspection robots

The exploitation of offshore subsea oil and gas wells arose in the early 1970’s
from the rapid inflation caused by the step increase in the price of oil from land
imposed by OPEC. So the need to inspect offshore oil risers arose at this time.
Manned submarines for sea bed and riser inspection were rapidly followed by
unmanned i.e. robot submarines. In 1976, the SONAMARINE robot underwent
initial trials in a deep water lake at Stoney Cove, Leicestershire. The robot had a
multi-axis manipulator to handle an ultrasound probe and the umbilical down
which the ultrasound signals were sent and returned exceeded 600m, permitting
operation in the deepest parts of the North Sea gas fields. The author was in
charge of the ultrasonic data collection and to his knowledge, was the first to
report robotic ultrasonic NDT in deep water.
 27

3. Further Developments

Wheeled climbing vehicle technology was further adapted to the inspection of storage
tanks and ship hulls involving greater climbing heights and thus greater payload
requirements because of the longer umbilical service cables required. DNV, RTD and
Silver Wing are examples of EU companies that pioneered this technology. Traditionally
magnetic wheels were used and the vehicles were prone to slip on dirty or greasy
surfaces. Thus use of electromagnets separate from the wheels reduced the slipping
problem but led to heavier and bulky robots. The latest state of the art is represented by
the OCTOPUS robot of Cybemetix which uses powerful magnetic plates to generate
more than 600Kg of contact pressure on the climbing surface, benefiting from the latest
improvements in rare earth magnet technology. Originally designed for ship hull
cleaning it has recently been fitted with ultrasound probes for hull inspection.

In later years, probably the first use of a climbing vehicle for inspection of ship hulls
dates back to the early 1970’s when a robot was custom designed to ultrasonically
inspect the Polaris Nuclear Submarines in dry Dock. An intrinsically safe system was
specified and this gave rise to a climbing vehicle using pneumatics both for propulsion
and foot suction. In the late 1980’s Portech developed this technology further for
cleaninghnspection work in nuclear power plant.

4. State of the Art

The rapid increase in computer processing power in the last 15 years has allowed a
commensurate increase in sophistication in automated and robotic NDT. However there
was and still is a tendency in the NDT industry for automated systems to be custom
designed too carry out just one task. One example is complex manipulators for
inspecting a specific nodal joint in a specific design of coolant circuit in a nuclear power
plant. After a single joint inspection at one outage the expensive system would lie in a
cupboard somewhere until the next outage at the power plant. Therefore in the early
1990’s much research has been carried out, especially under European funded
programmes, in attempting to find more generic solutions to robotic NDT. The Research
Centre in Automated and Robotic NDT (RCART) at London South Bank University
(LSBU) were the first to report a climbing robot which could climb to 30 metres carrying
a 6 axis PUMA arm, able to scan an ultrasound probe through an continuously variable
set of scanning routines, thus allowing multiple task NDT with one system. Later this
work was extended to include a 7 - axis arm designed by Ansaldo in the late 1990’s
which at 22kg, at the time was the world’s lightest arm available for NDT.

At RCART since 1992 Climbing and swimming robot NDT systems have been
prototyped to cover all the key demand areas for remote automated NDT, namely (i)
aircraft fuselages and wings, (ii) nodal joint inspection in nuclear power plant in a
radioactive environment, (ii) container ship hulls, (iii) storage tank floors whilst still full
of oil (iv) inside nuclear pressure vessels full of radioactive coolant . Key members of
these partnerships have been TWI, and Zenon in the projects RRNDT,
28

ROBTANKINSPEC, ROBAIR FPSOINSPECT and RIMIMI. The ROBTANK vehicle

for tank floor inspection and wall inspection under oil won the CLAWAR 2004 prize for
industrial innovation. In PIPESCAN, ZENON designed a novel inspection robot that
could walk and climb round pipes with thick protection thermal lagging, carrying out not
contact inspection underneath

In these projects interesting solutions were implemented in making robots that could
change surfaces under liquid or could climb round curved surfaces of very small radius
whilst deploying a heavy payload of scanned sensors in a stable manner. A number of
mechanisms involving flexible feet, ankle and thigh joints were successfully developed
to transport these heavy loads.

5. TheFuture

Looking to the future RCART, with ZENON , CYBERNETIX, BISSIACH and CARRU,
NAFTOSOL and TERNA are protoyping a climbing robot cell of three cooperating
robots for simultaneous welding and weld NDT in ship storage tan and wind turbine
manufacture (Figure 1). This project takes the world of robotic manufacturing out of the
factory floor and into the field. With TWI , RCART are also involved in the prototyping
of a robot for in situ inspection of turbine blades in offshore wind farms using the novel
technique of nanofocal computed axial Xray tomography (Figure 2).

6. Conclusions

Nondestructive testing is an industrial area in which climbing and walking robots can
truly be said to have made a practical impact. This impact continues on a rapidly
climbing curve. A particularly likely development is to extend the concept of networked
robots carrying NDT tools and other working tools to smaller and smaller robots so that
operation in highly confined spaces such as arise in nuclear power stations can be camed
out.
 29

.
Figure 1. The CROCELLS project (Climbing robot cells) By distributing a
payload of working tools over several climbing and cooperating robots it is possible
to carry out complex manufacturing tasks in situ on large structures such as ship
hulls, storage tanks, wind turbines and over civil engineering structures. The
distributed payload means that the robots can be kept small so that they can climb
on highly curved surfaces. In the CROCELLS project real time NDT of long weld
runs will be carried out on long weld runs simultaneously with the welding. So the
NDT data can be used to control the weld process parameters in real time and
instant weld repairs carried out if the NDT data indicate substandard welding
30

Figure 2 The CONCEPT project: nanofocal compute X ray to~ography for

three dimensional imaging of defects in fibre reinforced wind turbine blades. A
ring shaped climbing robot with magnet adhesion over a large area
provides the necessary support for a heavy payload with large overturning
moments. The payload consists of an x ray generator and detector array
scanned in x and y and with angular scans provided by blade rotation to
achieve the 31)inage reconstruction
 31

Figure 3. As Figure 2 with an enlarged view of the climbing ring robot and
scanner
32

Figure 4. As figure 3 but the ring climbing robot is replaced by a smaller

robot with less payload capbability which still adapts to the circMmference
of the turbine tower. The final choice between a ring design and the smaller
robot will depend on the final design payload presented by the scanner
 BIOMECHANICS AND ROBOTICS

NEVILLE HOGANt
Department of Mechanical Engineering,
Department of Brain and Cognitive Sciences
Newman Laboratory for Biomechanics and Human Rehabilitation
Massachusetts Institute of Technology, USA

1. Introduction

Robots are well on their way to becoming commonplace domestic appliances

but to realize their true potential requires the perfection of contact robotics,
machines that physically cooperate with humans. One pioneering application
requiring close physical cooperation between robots and humans is the delivery
of physiotherapy to facilitate recovery after orthopedic and especially
neurological injury. However, controlling robots to interact effectively with
humans presents unique challenges. A quantitative knowledge of human motor
and sensory performance is important to optimize machines for human contact. I
will briefly review our success with robotic treatment of upper-limb motor
disorders and how it has profited from the availability of relatively detailed
quantitative knowledge of unimpaired biomechanics and neural control of arm
movement.

Robotic treatment of lower-limb motor disorders presents unique challenges as

the control of locomotion is substantially more difficult and less well understood.
On the engineering side, locomotion presents far greater challenges than arm
movement. The dynamic process is fundamentally nonlinear, including
prominent discontinuities. It is a hybrid dynamic system, switching between
regimes governed by different dynamic equations. Add to these the paramount
challenge of safely managing physical contact and interaction with frail human
subjects. On the biological side, the relative importance of centrally-specified
limb trajectories and semi-autonomous peripheral neural oscillators such as the
putative locomotor pattern generator remains unclear. Furthermore, while
dynamics is largely subordinated to kinematic goals in upper-limb motor

nc? illc(?iiinit.cdu

33
34

behavior, musculo-skeletal mechanics appears to play a central role in

controlling the lower limbs.

However, the relative importance of musculo-skeletal mechanics suggests that

minimal actuation may provide substantial control authority; motors at every
joint are almost certainly unnecessary. While this has obvious advantages for
minimally-encumbering machine design, it makes the control problem even
more difficult as the system is underactuated. I will discuss our initial
investigations of the feasibility of controlling locomotion and assisting its
recovery using distal actuation, in particular a two degree-of-freedom robot
interacting with the ankle.
Advances in Climbing Robots
This page intentionally left blank
 A CPG WITH FORCE FEEDBACK FOR A STATICALLY
STABLE QUADRUPED GAIT*

JOSE CAPPELLETTO, PABLO ESTEVEZ, GERARDO FERNANDEZ-LOPEZ AND

JUAN C. GRIECO
Mechatronics Group. Electronics Department, Sim6n Bolivar University,
Sartenejas. Caracas 1020, Venezuela.

Abstract. In this paper it is described the design and test of a system for statically stable
gait generation for legged robots using a Central Pattern Generator (CPG) and normal
force feedback. The CPG, based on Amplitude Controlled Phase Oscillator (ACPO),
coordinates the generation of temporal references for each robot leg. The ACPO phases
are passed through a parametric function which transforms them into the desired
rectangular leg trajectories. Another parametric function generates a normalized force
reference for each leg, which is employed to compute three different desired Center of
Gravity (COG). Those references were used to enhance platform static stability margin
(SSM), by the addition of a simple PI controller and ground contact forces measurements.
Simulations were performed using a dynamic robot simulator (Webots) for a quadruped
and hexapod platform, and a real quadruped platform with 3-DOF per leg was also used.
Simulated and experimental results obtained for different controlled environments are
shown, and analyzed in order to identify the stability improvement of the CPG gait
generator by the addition of normalized contact force feedback to the spatial reference
generator.
Keywords: Central Pattern Generator, Amplitude Controlled Phase Oscillators, Force
Feedback, Static Stability, Legged Robots.

1. Introduction
Recent works in legged robots includes biologically inspired concepts as the
Central Pattern Generator (CPG). In animals, this system is able to generate
coordinate articulations movements based on local feedback and high level
instructions. Several authors have used recurrent neural networks to act as CPG,
behaving as dynamic oscillators to control leg movements (CTRNN) [l].
Despite the ability of C T R " to model dynamic systems, the proposed training
methods based in Genetic Algorithms (GAS) and modified real time back
propagation does not provide an analytical method to synthesize desired gaits
for robots with different leg configurations. J. Buchli proposed to use Amplitude
Controlled Phase Oscillators (ACPO) to describe the dynamics in robot

* This work was supported in part by Simon Bolivar University Research & Development Deanship.

37
38

locomotion [2]. This approach provides a direct way to control temporal

coordination between robot legs. However, it remains unsolved the relation
between those phase references and spatial leg references. It has been proposed
the utilization of simple feed forward neural networks to perform nonlinear
spatial transformation between different temporal references into motor position
references, as described by Cappelletto et aZ[3]. By using feed forward neural
networks it is also possible to perform soft transitions between different leg
trajectories. It has been reported the appearance of irregularities on leg
trajectory due to over fitting during neural network training; those undesired
behaviors can reduce stability during robot walking.
The work described in this paper proposes a solution for the problem
previously described replacing the spatial transformation subsystem by a
parametric transformation into a controlled rectangular trajectory. A direct way
to control the leg support factor (p) through a companding curve is also added.
Using this approach it is possible to control the desired gait mode and the
desired leg trajectory in direct and separated way. It is included a static stability
enhancement subsystem based on geometrically generated references for the
center of gravity, thus improving the robot walking for some marginally stable
gaits.
This paper is organized as follows: Section 2 describes the general
architecture for the gait generator and force based stability controller; and
includes a detailed description of the spatial and temporal generator subsystems.
The three COP generators, the stability enhancement and controller are
presented in the Section 3. Experimental setups for the simulation and real
platform are shown in Section 4; Section 5 and 6 shows the obtained results and
analysis and the main paper conclusions.

2. General Architecture
The CPG is implemented through two separated subsystems, making possible
to cope with possible changes in the mechanical platform with minor
modifications in the system. The walking adaptation process can be performed
by acting in the desired subsystem, thus allowing synthesizing different gait
modes with the same CPG model. Figure I shows a general description of the
main subsystems present in the pattern generator model here described.
 39

Enhancement
I I

Fig. 1. General System Architecture

Each subsystem was implemented as follows:

2.1. T e m p ~ rreference
~I subsystem:
As stated before, this subsystem generates the temporal references according
to the desired gait mode. Each walking mode can be described as a set of phase
relations between each leg, which can be translated in a direct way as the
rotation angles for the coupling matrix between all the oscillators’ nodes (21. In
the solution here described, it is used only the phase of the state vector for each
node. By this way is possible to neglect amplitude perturbations due to abrupt
switching between phase references corresponding to different gait modes. For
the quadruped model it was employed a fully interconnected architecture, and
for the hexapod model it were used connections between neighbors legs. It was
discarded the use of full interconnections between all legs for the hexapod
model because the excess of redundant connections. The int~rconnectionsfor
each robot model are illustrated in Figure 2.
A companding curve is applied to each oscillator state vector, acting over the
phase according to the desired support factor (p). This technique is described
with more detail in 141.
40

FI11
0 0
ACPO -*------e ACPO FD/2

1 f:

Quadruped Hexapod
Fig. 2. ACPO connections for Quadruped and Hexapod

2.2. Spatial reference subsystem:

To transform the temporal references given by the phases of the output
vectors for the previous subsystems, it was employed a parametric description
of the desired spatial leg trajectory. The trajectory here employed correspond to
a rectangle described by its height (H) that controls the elevation of the leg
between support and transfer phase, the length (L) that is equivalent to the stride
length thus giving control over the platform speed. It was also included a
displacement factor to control the separation between the leg and the main body.
This trajectory is parameterized for the variable t (0 to l), and the resulting
space reference point (X, Y , Z) for the leg is converted into actuators space
through inverse kinematics.

3. Stability enhancement
The approach previously described generates coordinated trajectories for all
the legs of a walking platform, given the geometrical parameters for the leg
trajectories, and inter legs phase relationships. Nevertheless, these trajectories
do not ensure the static stability of the platform since no attention is paid to its
weight distribution. This results on the platform falling periodically as a
consequence of its center of gravity (COG) being projected out of the support
polygon generated by its legs. Also, irregularities in the terrain or changes in the
weight distribution of the platform may yield to static instability.
In order to increase the platform Static Stability Margin (SSM) a control loop
is included at the end of the spatial references generation system. The control
 41

action is to displace the robot main body from its original position, keeping its
COG inside of the support polygon and far from its borders, and thus increasing
the SSM according McGhee and Frank’s criterion [ 5 ] [ 6 ] .Such displacement is
done in a plane parallel to the support one.

3.1. CG reference generator

Based on the phase reference generator, a force reference generator system
produces a normalized variable indicating how much force is the leg ready to
bear on each part of its gait cycle. This variable has a positive value (100) when
the leg is in support phase and a 0 value for the transfer phase. A short slope is
generated between these values to allow soft transition when possible. Using
this force reference, three algorithms are utilized to generate an appropriate
COG reference for the stability enhancement system.
The first one, named Balanced Forces Point (BFP), calculates an average of
all supporting leg tips positions using their referential forces as weights (Eq. 1).
The legs on transfer phase are naturally ruled out because of their null force
reference, and the slopes in the force references allows for soft transitions
between changes of the BFP. The BFP is always located inside the convex hull
of the support polygon, and gives a balanced distribution of effort between the
legs.
4 :Referential Pressure in leg i.
(Xi,
Yi) :Position of the tip of leg i.

It is easy to obtain support legs distributions yielding to a location of the BFP

with sub-optimal SSM. Nevertheless, experiments show that for the kind of
support distributions usually found in legged platforms and for small number of
legs, the BFP shows and acceptable performance.
The second algorithm identifies the desired convex support polygon based on
the referential leg forces, and calculates its Area Centroid (AC) (Eq. 2). Given
the convexity of the support polygon, this point is also always inside of it. By
ensuring a balanced distribution of the support polygon area around itself, the
AC locates the reference at a balanced distance of the polygon borders.
42

PorVgon
The third algorithm tries to reduce the computing requirements of the
previous one, while taking into account the most important condition: distance
to the borders. The convex support polygon Contour Centroid (CC) is calculated
by averaging all points in the polygon contour (Eq. 3). It tends to balance the
distance to all its borders.

J
contour

3.2. Control loop

The control loop consist of a COG projection reference generated from the
trajectories and force references of each leg, a scale factor (sf) required to avoid
references outside the robots work area and a PI controller generating the
required horizontal displacements for the platform.

Desired
Legs
Trajectories

I
BFP
CG
CA -b Ref -b PI Quad rCG
cc

Fig. 3. Stability Enhancement Control Loop

The feedback is obtained from normal contact force sensors located at the tip
of each leg. Those forces are combined with leg positions to compute the
vertical projection of the COG into the support polygon given static equilibrium
conditions. Because the SSM enhancement can be described as a trajectory
tracking problem a PI controller was employed.
 43

4. Experimental setup
This section describes the three main tests applied to the system architecture
previously described. In the frst test were compared the 3 COG references
shown in Section 3.1, using spatial and temporal references that correspond to
valid walking movements in the quadruped and the hexapod. The second test
consisted in dynamic simulations of both robot models using WebotsTM under
different environments, including uneven terrain and uneven weight
distribution, with and without the control loop. Finally, several tests were
performed on the real platform in order to illustrate the impact of the force
based feedback on the SSM for this CPG-based model.

4.1. Reference tests

Once the base gaits were synthesized, simulations were made for the COG
reference generation algorithms in the quadruped and hexapod platforms with
different values of p, and the standard McGhee criterion was used to evaluate
them. This way, standard walking support polygons were utilized for the tested.

4.2. Simulations
To study with greater detail the system here described, a hexapod and a
quadruped platform, both with 3 DOF legs, were simulated using WebotsTM.
The dynamic model is similar to the real quadruped robot, built on acrylic and
polycarbonate, and metal-plastic actuators (servomotors). The simulated
quadruped dimensions were 240x190 mm, and 335xl50mm for the hexapod.
The CPG phase relations were configured to generate a crawling gait in the
hexapod robot, and the scaling factor for the COG position was set to 0.4 for
both coordinates (X-Y) to ensure leg positions within their working space. With
this setup, different p values were used to test the performance of the platform,
both with and without stability enhancement, and using the three proposed
methods. The vertical projection of the center of gravity of the robot was
measured using the force sensors information for all the tests where the platform
showed a statically stable behavior, and this data was combined with the
measurements of the real legs position to estimate the SSM.
An equivalent test was conducted on the simulated quadruped robot,
resembling the Lynxmotion quadruped platform.

4.3. Real quadruped platform

In this test it was employed the Lynxmotion Quadruped-3 with 3 degrees of
freedom per leg, with servo motors as actuators. The robot dimensions are
44

240x190 mm, and the maximum range for each leg is 133mm. The platform is
equipped with FlexiforceB sensors in a rotationally compliant base tied to the
tip of each leg, measuring the normal force between the support surface and the
robot. The sensor data was processed under a Win32 C/C++ based GUI that
performed real time filtering and platform control with a base operating rate of
50 Hz.
Several tests were done with the real platform in order to replicate some cases
evaluated with the simulated quadruped, including uneven weight distribution.

5. Results and Analysis

Proceeding with the experiments previously listed, it was obtained the
following simulated and real data.
Computing the SSM of the three COG references for different p values on the
hexapod platform, is produced the data shown in Table 1.

Table 1 McGhee Statistics for CG references on the hexapod platform.

A similar test was ran for the quadruped and the obtained data is shown in
Table 2.

Table 2 McGhee Statistics for CG references on the quadruped platform.

The data of Table 1 and Table 2 shows how the average SSM for all COG
references tends to increase with the value of p, as could be expected for larger
support polygons. The BFP method provide a fast way to compute the desired
COG, without involving convex hull computation, and gives SSM pretty close
 45

to those obtained with AC and CC for quadruped robot. However, for the
hexapod platform it doesn’t show any important improvement and reaches
moderate SSM compared to those obtained with AC and CC: 21.2% and 19.5%
lower average SSM for all p.
Testing the whole system with the PI controller for several setups, the
resulting SSM are shown in Table 3. The controller constants were P=0.6 and
I=0.05, and sf=0.3. The step length was L=95mm, and p=0.85 for crawling
mode. For the real quadruped the controller constants were P = 0.3 and I = 0.08.

Table 3. Average measured SSM for different experimental and simulated setups

It can be observed the similar response for the three COG for all the cases.
The addition of the control loop increased the average SSM for the performed
tests, compared to the performance of the platform without stability
enhancement. For the last three tests, that included perturbations like uneven
weight distribution or ground slope, the simulated and the real platforms weren’t
able to walk in a stable way with the stability control disabled.

6. Conclusions

Based on the experimental results previously described, it was shown that the
proposed CPG architecture is able to produce statically stable gaits for
quadruped and hexapod robots. The decomposition of the walking problem into
separated subsystems simplifies the adaptation of the CPG for different gait
modes, and kinematical robot architecture by modifying only a specific
component of the model while keeping the others unmodified.
46

All the three COG generation algorithms provided references that increase
the platform SSM under different operative conditions. It was also observed that
for higher support factors the SSM also increased.
The BFP-COP generation method provides a soft transition between COG
references positions, however it exhibits a SSM lower than those obtained for
the other two geometrically based methods: Area Centroid and Contour
Centroid.

References and Bibliography

1. J. Hillel, Randall D. Beer and J. C. Gallagher. “Evolution and Analysis of
Model CPGs for Walking: I. Dynamical Modules”. Journal of
Computational Neuroscience. pp 99- 118, 1999.
2. J. Buchli and A. J. Ijspeert, “Distributed Central Pattern Generator Model
for Robotics Application Based on Phase Sensitivity Analysis”, In Proc. Of
1“ Intl. Workshop Bio-ADIT. 2004
3. J. Cappelletto, P. EstBvez, W. Medina, L. Fermin, J. M. Bogado, J. C.
Grieco and G. Fernandez-L6pez. “Gait Synthesis and Modulation for
Quadruped Robot Locomotion using a Simple Feed-Forward Neural
Network”. Lecture Notes in Computer Science, Vol 4029, pp 73 1-739,
2006.
4. J. Cappelletto, P. Estevez, J. C. Grieco, G. Fernhndez-Lopez and M.
Armada, “A CPG-Based Model for Gait Synthesis in Legged Robot
Locomotion”. In Proceedings of CLAWARO6.pp 59-64,2006.
5. McGhee, R. B. and Frank, A. A. “On the stability properties of quadruped
creeping gaits”. Mathematical Bioscience, Vol. 3 , pp. 33 1-351, 1968.
6. E. Garcia, J. Estremera and P. Gonzalez Dos-Santos. “A Classification of
Stability Margins for Walking Robots”. 5th Int. ConJ Climbing and
Walking Robots, Paris, France, 25-27 Sept 2002.
7. Y. Fukuoka , H. Kimura and A. H. Cohen. “Adaptive Dynamic Walking of
a Quadruped Robot on Irregular Terrain Based on Biological Concepts”.
Intl. Journal of Robots. V0122, No 3-4, pp 187-202,2003.
8. H. Kimura, Y. Fukuoka, Y . Hada and K. Takase. “Three-Dimensional
Adaptive Dynamic Walking of a Quadruped - Rolling Motion Feedback to
CPGs Controlling Pitching Motion”. In Proc. of ICRA02, pp 2228-2233.
2002.
9. S. Hirose, H. Tsukagoshi and K. Yoneda. “Normalized Energy Stability
Margin and its Contour of Walking Vehicles on Rough Terrain”. In Proc.
ICRA 2001. Vol 1, pp 181-186,2001.
10. P. Gonzalez de Santos, J. Estremera, E. Garcia and M. Armada. “Including
Joint Torques and Power Consumption in the Stability Margin of Walking
Robots”. Autonomous Robots, Vol. 18, pp 43-57,2005.
 A Sliding Sock Locomotion Module for a Rescue Robot

Luca Rimassa, Matteo Zoppi and Rezia Molfino

University of Genoa, PMAR Robotics research group,
Genoa, 16145, Italy, www.dimec.unige.it/PMAR
E-mail: (rimassa;zoppi;molfino) 4dimec. unage. it

Rescue robotics for disaster response deserves practical importance. Refer-

ence tasks are inspection and monitoring of collapsed structures and disaster
sites with searching/localization of victims. Mobile robots using unconventional
locomotion strategies are developed and applied (crawling, peristaltic, serpen-
tine). The identification of effective locomotion mechanisms and especially their
practical realization is one major research issue; often reproducing effectively
bioinspired mechanisms is challenging with today's available technologies.
This paper presents a development step toward highly efficient crawling lo-
comotion: A continuous membrane sliding along the entire body of the module
provides distributed trust with minimum risk t o get stuck. At one extremity of
the module, the membrane rolls in, slides back and rolls out again at the other
extremity.
The design is discussed and a physical prototype is presented.

Keywords: Rescue Robot; Sliding Sock Locomotion; Serpentine Robot

1. Introduction
Mini and micro rescue robots are used in calamity response operations for
search and localization of victims in collapsed buildings and disaster sites.
Three classes of rubble environments have been identified based on
structure and size of voids available:' 0 semi-structured (partially fallen
buildings whose original layout is globally preserved); confined (fully col-
lapsed constructions with voids comparable to human size); 0 sub-human
confined (fully collapsed structures bring into very compact debris with
very tight voids).
Confined and sub-human confined spaces can be explored only us-
ing worm rescue robots. Robot reconfigurability to adapt t o mis-
sion/environment requirements can be obtained by modular architectures:
a suitable sequence - number and assortment - of modules with same or
different sets of functionalities (locomotion, steering, carrying sensors).

47
48

(4 (b)
Figure 1. Increase of PR with the number of traeks (a) and PR= 1 with flexible belt
track (b)

One main research issue is locomotion. Several methods have been pro-
posed to move effectively in rubble debris. The more the trusting action
is distributed around the body of the robot, the higher is the overall ad-
vancing capability of the robot and the lower the risk that it gets stuck
because some passive part of its body (not generating trust) touches the
environment.
The efficiency of a locomotion mechanism can be ranked based on a
propuls~onratio (PR)2 defined as the ratio between surface providing trust
and total boundary surface,
No rescue robots developed have propulsion ratio really close to one,
which is the ideal PR realized when all body boundary surface contributes
to locomotion. Some systems with the highest propulsion ratio are: Gembu3
using wheeled modules; Kohga4 composed of tracked locomotion modules
connected by passive and active joints; ACM' merging snakelike loco-
motion to traditional locomotion in wheeled joint modules; MOIRA6 and
O m n ~ t r e ~ dwith
2 tracks distributed around the body, making the robot
indifferent to roll over, common to serpentine robots over rugged terrain.
The research has moved in the direction of PR= 1 by progressively
increasing the number of tracks around the body of locomotion modules,
Fig. l(a). Such increase cannot truly lead to unit P R and the design com-
plexity becomes soon very high. 5-6 tracks appear a limit difficult to over-
come in small robots.
Design can be simpler using tracks with flexible belts, Fig. l(b). PR= 1
is achieved in principle but, especially with long modules, strips can overlap
and part of the frame can contact the environment, with practical reduction
of the PR.

2. Continuous membrane
The real step forward is the realization of a fully moving body envelope.
 49

This is challenging and only two development experiences can be found

in the very recent state of the art. The f ~ r m e r , more
~ , ~ at a conceptual
stage, addresses the actuation of an elastic toroid membrane with oblong
section by expanding and contracting transversal rings integrated in the
membrane. It is proven that locomotion can be generated but no materials
are available today to realize the rings. The second de~elopment,~ with a
physical prototype realized, uses a rolling donut with circular toroid section
and wire actuation. An oblong section with external surface cylindrical,
required to move effectively in debris, cannot be obtained.
With this research background, we have started a two fold activity aimed
at the development of two locomotion modules with PR= 1 and oblong
section for rescue robots: the former is with rigid frame and uses traditional
solutions for the actuation of the sliding membrane; the second, still under
conceptual development, will be compliant to adapt to the shape of the
channel where it will advance. The following presents the main design steps
and the final set up of the locomotion module with rigid frame, for which
a prototype has been realized.

3. Design issues with continuous membrane

The design of a continuous membrane track system with internal rigid frame
is challenging. Assessed that it is not reasonable to imagine that the mem-
brane material can split and recombine during functioning, then the mem-
brane divides two fully separate domains, inside and outside. Two main
problems are faced: 0 the connection of the group composed of membrane
and its internal frame to the notch frame of the module; and 0 the actuation
of the membrane. The two problems are strictly interrelated and they have
been faced together.
Actuation splits in two further subproblems: how to make the membrane
moving and were placing the actuators: provided that a moving mechanism
is identified, if it operates on the membrane from inside and actuators
are inside as well, then the problem is how to provide actuation power.
Locomotion by anything embedded in the membrane body was not taken
into account at this stage because of complexity.
The connection of the membrane-frame group to the notch frame is
necessary to transmit the traction force from the membrane to the rest of
the robot. Independently from the membrane moving mechanism adopted,
no material bonding is possible to transmit this traction force. Two ways can
be identified instead: a coupling that can generate bonding forces through
the membrane and geometric bonding with distributed contact forces. Both
50

nc Crown of rolle
gear

Ma

crown of msgnsts

(4 (b) (c)
Figure 2. Actuation with worm gear (a); rollers (b); spur gears (c)

directions have been investigated.

For the former, we have studied combinations of magnetic fields gener-
ating repulsion forces and a stable equilibrium configuration for the notch
frame inside the membrane-frame assembly. Strong Neodymium magnets
are cheap and can be used to generate the fields; the considered magnets
layouts are shown in Fig. 2. The major drawback of this solution is the
quantity of ferromagnetic dust, always present in soil and construction ma-
terial, that would enter and lock the locomotion module very soon in real
application.
Then we decided to realize a geometric coupling by suitably shaping
the extremities of the frame inside the membrane. The design, shown in
the detailed view of Fig. 3, is further explained in Section 5.

4. Alternative membrane actuation solutions

Once defined the way to join the notch frame to the group membrane-
frame, we move to the discussion of the alternative actuations considered.
As mentioned in the Introduction, for this locomotion module conventional
mechanical solutions are adopted to have low cost and development time.

4.1. Actuation with worm gears

A suitably shaped worm gear, Fig. 2(a), rotating about the longitudinal
axis of the module, on the surface of the notch frame, is in contact with
the membrane. Some interference between the worm of the gear and the
membrane (which is supposed compliant) allows the gear worm to pull the
membrane. Rounds at the extremities of the gear make gearing progressive.
The large contact area results in safe and distributed pulling pressure. The
direction of motion can be easily inverted by inverting the direction of
rotation of the worm gear.
 51

Main drawbacks are the stretching in the membrane caused by the fric-
tion forces tangent to the worm gear and the high counter torque between
internal and notch frames, which is not easy to transmit if we use a geo-
metric coupling with contact forces to join notch and internal frames and
this coupling has to satisfy the axisymmetric geometry of the module.
A counter rotating worm gear can be added to balance most of the
coupling torque but this does not solve or makes even worse the problem
of membrane stretching caused by worm gear friction forces.
To reduce both friction stretching and counter torque, the worm gear is
replaced by a crown of rolling elements, Fig. 2(b), with longitudinal crests
on the notch to balance the residual counter torque. Also this alternative
design has drawbacks: it is complex and the rolling elements have small
contact area to pull the membrane with risk of damage.

4.2. Actuation with s p u r gears

To overcome the drawbacks of worm gears, design with spur gears with
axes orthogonal to the longitudinal axis of the module have been developed,
Fig. 2(c). The teeth of each gear etch the membrane and pull it tangentially
with no friction torque generated. The continuous membrane in between the
gears distributes the localized pulling forces so that finally an equivalent
continuous traction of the membrane is realized. A long worm gear coaxial
to the notch frame can transmit motion to all spur gears together.
The main design issue is finding an effective disposition of the gears on
the notch frame.
The limitations on minimum radius and thickness of the gears not to
damage the membrane result in a limit on the maximum number of gears
that the notch frame can host (minimum angular spacing of gears at each
transversal section of the notch). A distributed pulling action is obtained
by more sets of gears at different transversal sections, so that an angular
offset can be introduced for a more homogeneous distribution of the pulling
forces along the membrane.

5. Final module design

The final design, Fig. 3, meets the advantages of spur gears and geometric
joining of membrane and notch frames.
The membrane frame is cylindrical with two inward lips at the sides.
The membrane is first winded and bonded to realize a tube; then the tube
is winded around the membrane frame and bonded again to get the closed
52

Figure 3. Final design: virtual model and pictures of the realized prototype

final topology.
After attentive evaluation of alternative elastomeric materials with suit-
able softness, compliance and mechanical and fretting resistance character-
istics, a polychlorophene micro foam is selected. This material is commonly
used for diving wetsuit and is very low cost.
The membrane slides along the frame. A low friction factor is obtained
by suitable selection of materials and injecting a film of lubricant inside the
closed membrane. For the membrane frame Teflon is used; a nylon fabric
sheet is glued to the internal side of the membrane with weaving directions
oriented along the sliding direction to further reduce friction. As lubricant
a silicon based fluid is adopted.
The notch frame, realized in PVC, embeds two sets of 5 gears at each
side. It is realized in 4 parts (two shells and two covers) plus gears, trans-
missions and motor. Two worm gears provide motion to the spur gears and
actuation is provided by an electric motor in the center of the notch frame.
When the spur gear rotates the notch moves in the desired direction of
motion up to come in contact with the lip, then it starts pulling the mem-
brane and the higher is the resistance of the membrane to slip (related to
the resistance to advance of the locomotion module in the environment)
the more the teeth of the spur gears penetrate in the membrane pressed
against the lip, so getting higher trust.
The cabling used to interface the different modules of the complete robot
(electric power and signal bus) move on one side of the motor and in the
space available between the spur gears.
Compliance and stiffness of the membrane allow large geometric toler-
ances on the notch module, whose manufacturing can be easy and cheap
(e.g. standard moulding). All is designed for manufacturing with rapid pro-
 53

totyping equipment. For minimum cost, gears, some axes and other small
components of the module are commercial L EG 0 parts.
The final size of the locomotion module is 110 mm length and 50 mm
diameter plus membrane. Mass is about 200 g.
At the sides of the locomotion module mechanical/signal interfaces can
be arranged for fast and easy combination of modules.

6. Combined robot architectures

The locomotion module presented can be used to assemble worm robots
with various architectures. Distributed steering for small bending radius of
the worm and high steering capability can be realized with short locomotion
modules and steering modules between every two of them.
A compact design variant also suitable for inspection of pipes and struc-
tured channels is with one crown of gears instead of two on the notch frame
and with the membrane frame internally rounded as if the two lips de-
scribed in Section 5 had moved and become tangent. This rounded shape
matches the profile of the gears with minimum longitudinal backlash when
the direction of motion is inverted.
Externally the membrane is straight and the module has oblong section
although shorter than in the version with more crowns of gears, as required
to generate effective locomotion.
Membrane pulling is not as well distributed as in the case of multiple
crowns of spur gears, but sliding friction is much lower due to the shorter
membrane frame.
Due to the compact shape, the motor cannot be integrated in the module
anymore. One motor for all modules is placed at one side of the robot and
a flexible shaft is used to transmit motion to each and every module.
For pipe inspection the steering modules can be passive and the robot
is guided along the pipe by the pipe walls. Smaller bending radii are then
achieved.
The same robot architecture with some redesign of the modules can
be adapted to intestinal inspection. The main change should be in the
membrane, to be replaced or covered so that the external surface is rough
and with good grip on the intestinal walls.

7. Conclusions
A locomotion mechanism to realize worm robot locomotion modules with
PR= 1 has been developed and is presented in detail. The concept of con-
54

tinuous sliding membrane is developed and it is proven that conventional

actuation solutions can be used t o move the membrane. Preliminary tests of
a physical prototype are confirming the effectiveness of the design proposed.
This work belongs t o a larger research frame whose final goal is a sliding
membrane rescue robot with compliant body.
Design of steering modules will be also involved: even if the PR of each
locomotion module is 1, conventional steering modules open lateral win-
dows when the robot steers, reducing the overall PR. The step beyond is
masking these windows by adopting one continuous membrane through all
locomotion units. Modularity is still there because still a chain of identical
locomotion and steering modules is used.
The same rethinking of the robot will be possible when the compliant
locomotion module will have been developed.

Bibliography
1. R. Murphy, IEEE Trans. On Systems, Man, And Cybernetics Part C: Appli-
cations And Reviews 34 (2004).
2. J. Borenstein, G. Granosik and M. Hansen, The omnitread serpentine robot
- design and field performance, in Proc. SPIE Defense and Security Conf.,
Unmanned Ground Vehicle Technology V I I , (Orlando, Florida, 2005).
3. H. Kimura and S. Hirose, IEEE/RSJ Int. Conf. on Intelligent Robots and
System 1 (2002).
4. T. Kamegawa, T. Yamas, H. Igarashit and F. Matsuno, Development of the
snake-like rescue robot kohga, in Proc. 2004 IEEE Int. Conf. on Robotlcs d
Automation, (New Orleans, LA, 2004).
5. M. Mori and S. Hirose, Development of active cord mechanism acm-r3 with
agile 3d mobility, in Proc. IEEE/RSJ Int. Conf. on Intelligent Robots and
Systems IROS, (Maui, Hawaii, 2001).
6. K. Osuka and H. Kitajima, Development of mobile inspection robot for rescue
activities: Moira, in Proc. ZOO3 IEEWRSI Int. Conf. on Intelligent Robots and
Systems, (Las Vegas, Nevada, 2003).
7. M. Ingram and D, Hong, Whole skin locomotion inspired by amoeboid motil-
ity mechanisms, in Proc. 29th ASME Mechanisms and Robotics Conf., (Long
Beach, California, USA, 2005).
8. M. Ingram and D. Hong, Mechanics of the whole skin locomotion mechanism
concentric solid tube model: The effects of geometry and friction on the ef-
ficiency and force transmission characteristics, in Proc. IDETC/CIE ASME
Int. Design Eng. Techn. Conf. B Computers and Information in Eng. Conf.,
(Philadelphia, Pennsylvania, USA, 2006).
9. P. Breedveld, Development of a rolling stent endoscope, in 2006 IEEE / RAS-
EMBS Int. Conf. on Biomedical Robotics and Biomechatronics, (Pisa, Italy,
2006).
 A WHEELED WALL-CLIMBING ROBOT WITH A CLIMBING
LEG*

YILI.FU, ZHIHAI.LI, HEJIN.YANG, SHUGUO.WANG

Robotics Institute, Harbin Institute of Technology
Harbin, Heilongjiang Province, China

In resent years, wall-climbing robots used for special tasks such as rescue, inspection,
surveillance and reconnaissance have been studied hotly. But the technologies used in the
previous wall-climbing robots cannot meet the needs of these challengeable missions
mainly in mobile capability. A new prototype of self-contained wall-climbing robot is
developed, which integrate the advantages of wheeled robots and that of legged robots. It
is composed of a base body and a mechanical leg with 3 DOF. The base body is a big flat
suction cup with three-wheeled locomotion mechanism in it, and there is a small flat
sucker in the end of the mechanical leg. It can achieve quick motion on wall surface, as
well as obstacle- spanning on wall surfaces and smooth wall-to-wall transitions. The new
designed chamber seal has simple structure and has steady and reliable performance.
Kinematics model of the robot is established to analyze the robot's motion on the wall.
Furthermore, the locomotion gait of the robot is discussed.

1. Introduction
Wall-climbing robots integrate the mobile technology and adhesion technology,
it can adhere to wall surface and move around carrying kinds of tools. Therefore
it is expected to replace human to execute some dangerous missions in
hazardous environments. Since now, wall-climbing robots have been developed
and used in many industries, for example, maintaining nuclear facilities,
inspection of stage tanks, cleaning high-rise building, painting, and so on [1][2].
In resent years, wall-climbing robots used for special tasks such as rescue,
inspection, surveillance, reconnaissance, counter terrorism and so on, has
become a new branch of wall-climbing robots. These new tasks demand the
robot have small volume, low weight, excellent mobility and maneuverability.
But the technologies used in the previous wall-climb robots cannot meet the
need of these new special tasks. They always have big body, high weight, and
low agility. So, new prototypes of self-contained wall-climbing robot need to be
developed to execute these special tasks [3][4].

This work is supported by National Natural Science Foundation of China (60675051)

55
56

Since now, wall-climbing robots developed for surveillance and

reconnaissance tasks mostly use vacuum adhesion mechanism to make the
robots sticking on wall surfaces, because this method can suit walls such as
ceramic tile, glass, cement and brick walls. Although there have been some little
wall-climbing robot prototypes developed, there is also some problems required
to be settled. For example, the wheeled wall-climbing robot has high moving
speed, but its obstacle-spanning capability is poor. In other side, the legged
wall-climbing robot can span obstacles on the walls, such as bars, barriers and
ledges, but its moving speed is low.
Here, a new type of wall-climbing robot is proposed, which integrates the
advantages of wheeled robots and legged robots. It can achieve quick motion as
well as obstacle- spanning on wall surfaces, and even smooth wall-to-wall
transition by applying integrated locomotion mechanism. Vacuum based
adhesion mechanism is used to adapt to versatile surfaces.

2. ~ e c h a n Design
~~a~
This novel wall-climbing robot is composed of two parts: a base body, which is
a big flat suction cup; and a 3-DOF mechanical climbing leg with a little suction
cup in the end, as shown in Fig.1. Design of the robot’s mechanical system
mainly includes adhesion mechanism, sealing mechanism, wheeIed locomotion
mechanism, and mechanical leg.

Fig. 1 Outlook of the novel wall-climbing robot

The base body takes on a cuboid frame. It contains adhesion mechanism,

sealing mechanism and wheeled locomotion mechanism. Thus, the base body
can be an integrated robot alone, Besides, it needs to carry batteries, controller,
sensors and other tools. And when the climbing leg is idle, it will be a load to
the base body. Therefore, the base body needs to have high payload capability.
 57

2.1. Adhesion mechanism

An excellent adhesion mechanism is the guarantee for the robot’s reliable
working on wall surfaces. Main limitations of selecting adhesion method are the
materials and state of wall surface. In order to enable the robot to adapt to
versatile surfaces, vacuum adhesion technique is chosen. Considering working
surfaces may be porous and bumpy, vacuum rotor package is selected since it
has more advantageous than micro-pump in this condition. The vacuum rotor
package consists of DC motor, impeller, and exhaust cowling.
When the robot sticks on the wall, a vacuum chamber is formed between
the base body and surface. The chamber and vacuum rotor package constitute
the vacuum adhesion mechanism. Under the high speed rotation of the impeller,
air in the chamber is pumped out via the exhaust cowling. A low air pressure
under atmospheric pressure is generated in the vacuum chamber, so that a force
due to the pressure difference presses the robot attached to wall surface reliably.

2.2. Sealing mechanism

1 ,fixing hole
PO

airtight fabric chamber

flexible stuffing

\-Pl
Fig.2. Sketch and Force analysis of the new vacuum chamber seal

The negative pressure generated in the vacuum chamber need to be maintained

by sealing the gap between suction cup and wall surfaces. Two sealing methods
are always used: inflated rubber tube skirt seal and flexible bristle seal. The
former which is always composed of some springs and an inflated rubber tube
filled with air has perfect performance, but the effect is correlative to the air
pressure in the inflated tube. The performance is not steady considering air leak.
In order to improve the stability of seal, a new sealing skirt is designed as
shown in Fig.2, which is composed of inner flexible stuffing, outer smooth
airtight fabric. It can realize perfect effect as well as inflated tube skirt seal,
furthermore it has other advantages-self-press capability. When the vacuum
motor works, air flows from the outside into the chamber via the gap between
seal skirt and surface. A low pressure area is generated in the gap. Therefore,
the flexible skirt can be pressed against wall surface under air pressure
automatically. Additionally, this novel seal has a simple and firm structure; its
stable parameters make it have stable performance.
58

2.3. ~ e e l e locomotion
d mechunism
In the past, wall-climbing robots with a single suction cup always use 2 or 4
wheeled drives, but 3-wheeled drive is rarely referred. Comparing with the
former two modes, 3-wheel drive is both simple and steady, thus 3-wheel drive
is adopted. As shown in Fig.3, the locomotion mechanism consists of two drive
wheels and a castor wheel. The two driving wheels are driven by DC motors
independently, which are controlled by differential principle. Tires m o ~ t i n gon
the driving wheels heighten friction coefficient, and enhance damp between
robot and surface. Wheeled drive guarantees the robot have high moving speed.

Fig.3. 3-wheeled locomotion mechanism

2.4. ~echuniculclimbing leg

In order to improve robot’s mobility, a climbing leg inspired by inchworm is
equipped to the base body. This climbing leg has 3 DOF. It is composed of two
rotational joints, a translational joint and a flat suction disc as shown in Fig.1.
The two rotational joints are driven by DC motor respectively, and the
translational joint is realized by rack and pinion mechanism driven by a DC
motor. The climbing leg can help the robot climb over obstacles since there is
always protrusion or gap on the building wall [ 5 ] . The robot can also imitate the
inchworm’s movement on the tree.
Weight of the robot is 9 kg, the base body is 300(~)X300(L)X150(H) in
mm. Fully extended it is 900 mrn long. The wheeled mechanism’s speed is
0-lOm/s; step size of the legged mechanism is 0-450 mm.

3. ~ n e m a t i c sanalysis
In order to facilitate the analysis and control of the robot’s movement,
kinematics model of the robot is established. Considering the wheeled
locomotion mode and legged locomotion mode works independently in different
conditions, kinematics models of the two modes are settled respectively. They
are illustrated as follow.
For wheeled locomotion mode, the base body coordinate is established, as
shown in Fig.6. In the world coordinates, the positio~orientationof the base
 59

body can be expressed by (x,y,$)T , where 4 and b2 denote the two driving
motors' rotational angle respectively, and i denotes transmission ratio. The
wheeled locomotion mechanism's kinematics equation is expressed as Eq. (1).

[i]
Fig. 4 Coordinates frames of the base body Fig. 5 Coordinates frames of the mechanical leg

= 5 J I(8, L ) [ i 1, -1/L
case case
J1(e,i)= [sin 6 sine]
1/L
(1)

For wheeled locomotion mode, the assignment of coordinate frames based

on the Denavit-Hartenberg (D-H) method is illustrated in Fig.5. The link
parameters are shown in Table 1, 4, d2 and +3 are joint variables. Transformation
matrix o< (Eq.2) expresses the small sucker's position/orientation in coordinate
0. On the other hand, matrix 'r, ( 3 r , = 0 ~ )- ' expresses the base body's
position/orientation in coordinate 3.

Table 1. Link coordinate parameters of the mechanical leg.

Joint e d a a
Joint1 4 0 a1 -90'
Joint2 0 d2 0 +90'
Joint3 $3 0 a3 +90'
r-C(4+B3) 0 S(B,+8,) alcosB,-d,sinB,-a3sin(Bl +@)I
s(B,+e,) 0 c(B,+B3) a, sin@,+ d , cos8, +a, cos(8, + 0 3 )
0 1 0 0
0 0 1 J
4. Force analysis on the wall
If a wall climbing robot can reliably move on vertical surfaces, it needs to fulfill
some force constraints. Static force analysis is used to analyze these constraints,
which is executed in two cases: wheeled motion mode and legged motion mode.
60

4.1. Force analysis of the wheeled locomotion mechanism

When the robot works in wheeled motion mode, sticking and driving of the
robot rely on the friction force between the drive wheels and wall surface. The
robot can move in any orientation under differential control. They can be
classified in two main categories-driving wheels underside and driving wheels
upside, as shown in Fig.6 a) and b).

a) Driving wheels underside b) Driving wheels upside

Fig. 6 Forces and torques of the robot at two conditions
As shown in Fig.6 a), forces acting on the robot include gravity, negative
pressure adhesion force, supporting force by surface, and friction forces
between wheels and surface, they are denoted by G, Fp,FNAand f N A respectively.
Here friction between the seal and surface is ignored. When the driving wheels
is underside, and wheels doesn't slip. We have,

Since there is no slipping, friction between wheels and surface is static

friction. plmax= 0.8 , p2max = 0.05 .So, the friction forces can be calculated by,
f A = f B = P I e F N A 5: Plrnax'F~~ f C = P 2 ' F N C 5 P2rnax'FNC
9 (4)
By Eq. (3) and Eq. (4), we have,

FNA =(FPal+G*h)/41,F N C =(FP*l-G*h)/21

= FNB
Eq. (5) indicates the relationship between negative pressure force and
general gravity. It can be rewritten as,

H is a function of p I and H . As shown in Fig.7, H increases along with p,

increasing and p2decreasing. And min(H,,,,) = 1.751 . Along with H decreasing,
Fp decreases. The needed negative pressure F ~in
' practice can be calculated by,
 61

F,' = HaG*[S]
Where, [s] is secure coefficient.
If the robot's upper wheel is attached to the surface, we have,
FNc = (FPd-G*h)l21> 0
It can be rewritten as, F, 2 Gohll
So, the needed negative pressure F,' can be calculated by,
F,' 2 max(Hhw,,slde
*G*[S], Geh 11) (7)
When the driving wheels are upside as shown in Fig.6 b), force analysis can
be executed by the same procedure. We can get that,
2f+(P, +&u,).h
ff"P31de = ,F,' 2 max(H,,pslde~G*[S],
G-h /I) (8)
(PI -Pz )"E
From Fig.7, it can be found that load capacity is higher when driving
wheels are downside than when they are upside. Therefore, load capacity should
be calculated by the analysis when the wheels are upside.

4.2. Force a n a ~ s i ofthe

s ciim~ingieg mechanism
The obstacle-spanning of the robot relies on the lifting capability of the small
sucker in the end of the mechanical leg. Therefore, force analysis should be
used to analyze the lifting capability.
As shown in Fig.8, when the small sucker keeps staying on the surface, it
bears such forces as negative pressure Fp, normal force N, friction force f ,and
load force Fy,,which includes gravity of base body and climbing leg. We have,

!"In P

FY

Fig. 8 Forces analysis of the small sucker

62

c x=0 , c [ N o+n*(L-I)]*dl- F,) = 0

cy=O , p * ~ [ N o + n - ( L - I ) ] * d l - F=yO (9)

~kf(B)=O,-~[No+n~(L-I)]~I~dl-F,~---Fy~Hy
L =O
2
By Eq. (9), we have, F, = F y p (10)

Considering U , I ,u,,,=, by Eq. (10) we have, F, I ,u=*F, (12)

If the small sucker is attached to surface, N~ 2 0, Eq. (1 1) can be rewritten as,

F, 2 FY*6Hyl L (13)
By Eq. (12) and Eq. (13), the load is given by,

Relationship among Fy, Hy, and AP is indicated by Fig.9. Pdirrrepresents the

vacuum degree AP . F,, increases along with AP increasing with the constant Hy
Load capacity of the small sucker becomes higher with AP increasing.

Fig. 9 Relationship between F', Hand Pdg

The specifications of the two adhesion mechanisms are shown in table 2.

Table 2. Specifications of the two adhesion mechanisms.

Vacuum degree Size (mm) Seal method Load capacity

Big sucker 9000-18000Pa 300(W) X300(L) 207-422N
of base body
~~~~~~

Small suck& Flexible Rubber 150-300N

15K-30Ka 150(W) X 540(L) seai
of climbing leg (H=I 50mm)
 63

5. Analysis and Simulation of robot’s movement on wall surfaces

The integrated locomotion mechanism improves the robot’s mobility evidently,
since it enable the robot to have two motion modes. The wheeled locomotion
mechanism is used to realize high speed movement. And legged mechanism is
used to span obstacles. Selection between the two modes is determined by the
condition of wall surface. The two motion modes are analyzed and simulated by
establishing 3D model and executing kinematics and dynamics analysis.
If the wall surface is smooth, the wheeled mechanism works. The adhesion
force press the robot against the surface steadily, actuators are controlled in a
differential way to drive the 3-wheeled locomotion mechanism to realize robot’s
omni-directional movement. As shown in Fig.10, the robot can move upwards,
downwards and along arbitrary directions in normal or inverse orientations.

Fig. 10 Robot in wheeled motion mode

When the robot is faced with obstacles in its routine, the mechanical leg is
controlled to realize obstacle-spanning movement. In this mode, the Fig.10

5) 6) 7) 8)
Fig. 11 Locomotion sequence to span a ledge on wall surface

1. The robot gets close to the ledge. Then base body sticks on the surface
f m l y and keep still, as shown in Fig. 11.1).
2. Extent the mechanical leg and drive it over the ledge. Then align the small
sucker in the end of the mechanical leg with wall surface and move it until a
secure grip is verified with micro air pumps enabled, as shown in Fig. 11.2).
3. Base body released from the wall, as shown in Fig. 1 1.3).
4. The mechanical leg lifts the base body over the ledge, as shown in Fig. 11.4).
5 . When base body lands on surface reliably, air pump works. After verifying
the base body attached to the surface securely, the small sucker released, as
shown in Fig. 11.5) and I 1h).
64

6. The mechanical leg is retracted and placed on the back of the base body,
as shown in Fig. 11.7) and 11.8). Obstacles-spanning movement is finished.
The above sequences are assumed to be realized by the on-board controller
autonomously. When the predefined strategy does not work, each joint and
sucker may also be controlled manually to assist the robot to achieve secure
movement. Besides, the mechanical climbing leg enables the robot has the
ability to transit between two walls, wall and ground or between wall and
ceiling. The alternate use of the two locomotion modes enables the robot to be
adapted to more surfaces.

6. Conclusion
This paper presents a novel wall-climbing robot, which integrate two
locomotion modes. This principle overcomes the limitations of previous wall-
climbing robots, and this robot prototype is better in term of capability, mobility,
and adaptability than previous ones. Integration of the wheeled locomotion
mode and the legged-climbing locomotion mode evidently improves the
mobility of the robot, so that the robot has both high moving speed and
excellent obstacle-spanning capability. The performance makes the robot could
be competent for versatile missions. The next step is to design embedded
controller board and build sensory system to realize a self-contained robot
system. Based on these, autonomous capability can be implemented to improve
the performance of the robot.

References
1. A. Nishi, Development of wall-climbing robots, Computer and Electrical
Engineering, 22(2): 123 (1996).
2. Y .Wang, S .L.Liu, D.G.Xu, Y .Z.Zhao, H.Shao and X.S.Gao, Development
& Application of Wall-Climbing Robots, Proceedings of the 1999 IEEE
International Conference on Robotics & Automation, 1207 (1999).
3. J. Xiao, J.Z. Xiao, N. Xi, R.Lal Tummala, and Ranjan Mukherjee, Fuzzy
Controller for Wall-Climbing Microrobots, IEEE TRANSACTIONS ON
FUZZY SYSTEMS, VOL. 12, NO. 4,466 (2004).
4. J. Z. Xiao, A. Sadegh, M. Elliott, A. Calle, A. Persad and H. M. Chiu,
Design of Mobile Robots with Wall Climbing Capability, Proceedings of
the 2005 IEEE/ASME International Conference on Advanced Intelligent
Mechatronics, 438 (2005).
5. B.L luk, A.A Collie and J.Billingsley, ROBUG 11: AN INTELLIGENT
WALL CLIMBING ROBOT, Proceedings of the 1991 IEEE International
Conference on Robotics and Automation, 2342 (1 99 1).
 AN EVOLVED NEURAL NETWORK FOR FAST
QUADRUPEDAL LOCOMOTION

IRENE MARKELIC
Bernstein Center for Computational Neuroscience, Bunsenstrasse I0
Goettingen,, 3 7073, Germany

KEYAN ZAHEDI
Fraunhofer Institute for Autonomous Intelligent Systems (AIS), Schloss Sankt Augustin
Sankt Augustin, 53757,Germany

This paper presents a modular neural network controller for fast locomotion of a
quadruped robot. It was generated by artificial evolution techniques using a physical
simulation of the Sony Aibo ERS-7. Co-evolution was used to develop neuromodules
controlling the single legs as well as the coordination between the four legs. The final
neurocontroller utilizes a central pattern generator and does not make use of available
sensory inputs. In experiments with the physical robot a top walking speed of 47.34 c d s
was measured, where lateral leg movement contributed considerably to the achieved high
velocity.

1. Introduction
The control of legged locomotion in robots is still a challenging
problem, especially for fast locomotion. Apart from some
exceptions like Raibert's robots [l] most legged machines move
rather slowly [2]. It has been found that locomotion in many
organisms is mostly driven by central pattern generators (CPGs)
[3]. These are neural networks producing a rhythmic pattern
without the need of sensory feedback [4].For walking machines a
CPG can be realized by an oscillatory artificial neural network.
Such oscillators were used for instance by Billard and Ijispeert to
realize a continuous passage between a walking, scratching and
lying down behavior of an Aibo robot [ 5 ] . Kimura used neural
oscillators to realize dynamic walking and running in a quadruped
[ 6 ] . To find appropriate controllers for a hexapod, Beer used
genetic algorithms and a simulation of the robot [7]. Finally he
transferred the generated controllers to a physical machine [S].

65
66

Cruse et al. developed neural networks by means of a simulation

and implemented them in a hexapod [9]. In this paper a recurrent
neural network for fast quadrupedal walking of a Sony Aibo ERS-
7 robot is presented. It was derived by artificial evolution
techniques and is for use in the international competition of robot
soccer. Since velocity is a key feature, many teams are concerned
with fast locomotion. Published velocities of Aibo walking using
the ERS-7 model are between 39.7 - 43.0 c d s [10,11,12,13]. The
paper is organized as follows: In section 2 (The Experimental
Setup) the neuron model is introduced, and the tools employed for
evolution and composition of the network are explained. In the
third section (Results) the resulting controller is presented and
analyzed. In the final section (Discussion) the results are discussed
and related to other works.

2. The Experimental Setup

All experiments were conducted using the time discrete dynamics
of networks with standard additive neurons, using the hyperbolic
tangent as transfer function ; i.e.,
N

a,(t + 1) = oi+ CWva(a,(t)), i=I,..N

j=1

where ai denotes the activation, Oi the bias term of the ith neuron, t
denotes a discrete time step, 00 refers to the weight of synapse
rom neuronj to i, and N is the total number of neurons in a given
network. Input neurons were treated as buffers for sensor signals.
For evolution the program package ISEE (Integrated Structure
Evolution Environment ) was used, which was designed to realize
Artificial Evolution experiments. It implements the E N 9
algorithm [14] and offers the physical simulator YARS based on
ODE (Open Dynamics Engine). ISEE permits to influence the
evolutionary process and the network by various parameters that
can be set during runtime, for example, to foster small networks by
assigning cost terms to neurons and synapses. It also supports co-
evolution, i.e. the simultaneous evolution of several populations.
Furthermore, the generation of network structure as well as
parameter optimization is possible.
 67

Figure 1.Left: The simulated robot. Right: The neurocontroller generating the fastest walking
behaviour. Circles denote neurons and are connected via synapses. Neurons C1 to C4 are connector
neurons which form the junction between leg modules and coordination module.

A dog-like robot, the Aibo ERS-7 was used as physical platform

for the evolved controller. Each of its four legs has three degrees of
freedom realized by the three motors, Ml, M2 and M3. The first
motor, M1, moves the leg back and forth, whereas M2 moves it
sideways. M3 moves only the lower limb of the leg back and forth.
For each of these motors a sensor (S 1, S2, and S3), is detecting its
deflection angle.
The physical simulation of the robot, as shown in Fig. 1 left, was
realized with YARS. As relevant features to be modelled for a
walking behavior were considered: body parts (head, trunk and
limbs) defined by their dimensions and weights as well as the 12
leg motors which were each defined by a deflection angle,
strength, and velocity, and finally the according 12 deflection
angle sensors. Since hinge torque and velocity of the motors used
for the ERS-7 are not published they had to be measured
experimentally. For maximal force 0.43 Ncm were used, and 286
"Is for maximal angular velocity. To instantiate the controller in
the robot it was rewritten as C++ program and incorporated into
the freely available German Team Software which provides access
to motor control and sensory information.
Based on the hypothesis that control of insect walking can be
considered hierarchical and modular [16], and assuming that both
fore legs and both hind legs each carry out the same step cycle, a
modular approach, and thus co-evolution, was chosen as an
appropriate method. It was decided to use one network per leg,
68

where the networks for both fore- and both hind legs were
identical. All leg modules were then connected to a coordination
module. A minimal leg controller contained seven neurons, one
input neuron per sensor and one output neuron per motor. The
seventh neuron was a hidden neuron by which the connection to
the coordination module was realized and which is referred to as
connector neuron. The resulting network formed the
neurocontroller, comp. Fig. 1, right.
When a reasonable walking behavior was observed during
simulation the corresponding controller was transferred to the
hardware, and the speed of the resulting gait measured. Therefore a
test course was set up with two labels, indicating one meter and the
robot had to cover this distance from a flying start and with fully
charged batteries while being filmed. Each neurocontroller was
tested 10 times, and its speed was set to be the average value of
these runs.
Various experiments were conducted, and the best results were
achieved when using a CPG as coordination module. The
employed CPG, depicted in Fig. 2A, was adopted from [17]. Its
output are four sinusoidal signals with equal phase shift (see Fig.
2, bottom left), of which each leg module received one in the
order: right hind leg, right foreleg, left hind leg and left foreleg.
This, together with the corresponding phase shift, generated a
reasonable walking gait. An evolution run was started by providing
initial network structures for the leg modules in order to avoid the
bootstrap-problem. They caused a pendulum like movement of the
legs which was triggered by sensory input. Parameter settings of
the evolutionary ENS3program enabled an alteration of structure,
bias and synaptic weights of these networks. Structure and
synaptic weights of the CPG were chosen to be fixed, because
already small changes often had undesirable effects for the output
signals, e.g. the loss of periodicity. The fitness function F w a s
given by
T 2

t=O

where t denotes discrete time, and x the position along the x-axis
of the simulated robot in world coordinates, which was equal to the
 69

covered forward distance. It was squared to increase selection

pressure. To punish individuals with high frequency oscillating
legs, a problem that frequently occurred, the change of direction in
leg movement was counted, denoted by c. It was then multiplied
by a factor, a, which was adjustable during runtime. In addition,
low obstacles were built into the simulated environment which
could not be overcome by individuals with fast oscillating leg
movements, so that they could not gain high fitness values. After
4320 generations a neurocontroller leading to a speed of 39.56
c d s was generated.

3. Results
The resulting network structure is shown in Fig.2, right. It is
observable that no further hidden neurons, apart from the
connector neuron, were part of the leg networks and that the only
driving force of the controller was the CPG. Although sensory
input from 12 sensors was available during evolution, only one
sensor in the hind leg modules was effciently used. It was found
that this sensor input was important for the initial behavior of the
simulated robot. It modulated the trajectory of the signals that
controlled M1 of both hind legs in such a way that the simulated
robot could get in a position from which it could start walking.
Because the physical robot was not started from such a position ( it
was put on the floor when moving) it was not affected by this
mechanism. Thus, a controller for walking without any
incorporation of sensory feedback had been evolved.
Signals were fed back into the CPG coming from the connector
neurons. Although the according synapse weights were small,
between [-0.08, 0.081 they had impact on the CPG's output signals,
as can be observed when comparing Fig. 2, bottom middle and
bottom right. Changes in amplitude (Cl, C4), period (Cl, C2, C3,
C4), phase shift (C4, C3) and signal shape (C3) are observable.
With this controller the robot did not walk in a straight line, which
was due to the different signals the motors received. Therefore
synaptic weights were carefully adjusted so that all legs received a
similar signal (see Fig. 2 top, right).
70

Figure 2. Top row: The CPG with connector neuron of each leg (CI-C4). The originally used CPG,
left, the CPG with evolutionary altered parameters, middle, and the manually adjusted CPG, right.
The corresponding bias values are listed in table 1. Bottom row: The plot below each CPG shows
the output of the four connector neurons. Each curve has is labelled with the correspondingneuron.

Furthermore parameters were altered to find out which stride

frequency and length were most appropriate. This was done
manually, but could have been done using evolution. As was
already indicated by the longer period time of CPG signals after
evolution, the best results were attained with a lower CPG signal
frequency of 1.8Hz. The initial frequency in comparison was
2.2Hz. In Fig. 3 the signal of M1 of the right foreleg is plotted
along with the according sensor signal. It is observable that the
motor lags behind the given signal and never reaches the target
amplitude. Thus, with a larger period the motor can follow the
given signal better, which finally results in a larger step length. In
contrast, a higher frequency automatically led to smaller steps and
due to the inertia of the motors to a lower speed of the walking
behavior.
Leg posture was determined by the bias values of the according
output neurons. Changing these values enabled a straighter posture
of the legs which caused larger steps. The final neurocontroller
(see Fig 2, right) generated a walking behavior with a speed of
45.95 c d s , with a top speed of 47.34 c d s . Videos of the
behaviour are available under
www.fraunhofer.de/-zahedi/aibo.html.The obtained gait is a walk
with duty factor p = 0.58 f 0.2 for all legs. Where duty factor is
the ratio between the duration that a leg has ground contact and the
 71

duration when it is lifted off the ground during one stride. One
major difference to other Aibo walking behaviors was the
incorporation of M2 in the hind leg modules, which resulted in a
sideways movement of these legs. It caused the robot to rock from
one side to the other while walking (compare video). Lesion
experiments showed that this resulted in a gain of speed of 1 1.8%.

Table 1. A comparison of bias values of the different CPGs. The

enumeration of neurons is according Fig. 2,nn=neuron number, bo
=bias of original CPG, ba=bias of evolutionary altered CPG, and
bop=bias of optimized CPG.
nn bo ba bop nn bo ba bop
C1 0 -0.003 0 5 0.01 0.017 0.0175 Fig. 3. Output o f M l and S1.
C2 0 0.0016 0 6 0 0.0053 0.0053 shows the actual
deflection of the motor.
c3 0 -0.002 0 7 0 0.0037 0.0037 mereasM1 can be
C4 0 0.0065 0 I8 0 -0.0044 -0.0044 interpreted as a setpoint
signal.

4. Discussion
Modular neuocontrollers were evolved to control a physical Aibo-
ERS-7 robot. Using a physical simulator and additional
adjustments, the final controller was able to generate a fast straight
walking behavior for this robot of about 47,3 c d s . It was found
that a sideway movement of hind legs resulted in a higher walking
velocity. This can be explained by the shift of the robot's center of
gravity from side to side, caused by the lateral hind leg movement.
When one bodyhalf was tilted towards the ground, carrying most
of the weight, the other side was moved upwards, leaving more
time and force for legs on this side to be moved.Thus, it is
concluded that a pace gait in addition to the horizontal body tilt
can lead to a higher walking speed for this particular robot, than a
trot or a walk Experiments by Beer suggest that sensory input is
always incorporated when available [8]. That this does not happen
in this case may result from the fact that sensory input was simply
not needed for the task at hand, because of the driving CPG.
Evolution took place in an environment that indeed contained
obstacles, but obstacle height was relatively low. Once it was
provided that legs were lifted up high enough there was no reason
for a more sophisticated behavior. That no sensory input was fed
72

into the network supports an hypothesis by Full and Koditschek

[ 191 that rapid locomotion relies rather on feedfonvard control than
on continuous sensorimotor feedback. This is thought to be due to
the little time available to process proprioceptive input [ 191. For
future work, it is desirable to steer the walking direction of the
robot. For this, visual sensory input could be used as a tropism
(e.g. using the robot's camera to follow a colored ball).

References
1. M.H. Raibert & I.E. Sutherland, Sci. Am., 248,44 (1983).
2. R.M. Alexander, Phys. Rev., 69, 1199 (1989).
3. S. Grillner, Curr. Opin. Neurobiol, 1(4), 231, (1991).
4. F. Delcomyn, Science, 210(4469), 492 (1980).
5. A. Billard & A.J. Ijspeert, Proc. IJCNN, VI, 637 (2000)
6. H. Kimura et al, Aut. Rob., 7(3), 247 (1999).
7. R.D. Beer & J.C. Gallagher, Adapt. Behav., 1, 91 (1992)
8. J.C. Gallagher et al., Robot. Auton. Syst., 19, 95 (1996).
9. H. Cruse et al., Adap. Behav., 3(4), 385 (1995).
10. A. Shinohara & A. Ishino, www.jollypochie.org/papers/
JollyPochie2004TR.pdf, (accessed 2006).
11. M. Quinlan et al., www.robots.newcastle.edu.au/
publicationns/NUbotFina1Report2005.pdf,(accessed 2006).
12. D. Cohen et al., www.cis.upenn.edu/robocup/UPennOl.pdJ;
(accessed 2006).
13. T. Rofer, 8thInternational Workshop on RoboCup, (2004).
14. F. Pasemann et al., Theory Biosc., 120,311 (2001).
15. M. Fujita & H. Kitano, Aut. Rob., 5, 7 (1998).
16. F. Delcomyn, Aut. Rob., 7,259 (1999).
17. B. Klaassen et al., VDI-Berichte, 1841,633 (2004).
18. F. Pasemann et al., Proc. IWANN, 2686,144 (2003).
19. R. Koditscheck, J. Exp. Biol, 22,3325 (1999).
Autonomous Climbing Motions for Connected Crawler Robots
Sho YOKOTA*, Yasuhiro OHYAMA,
Hiroshi HASHIMOTO and Jin-Hua SHE
School of Bionics, Tokyo University of Technology,
Tokyo, Japan
*E-mail: yokota@bs.teu.ac.jp

Kuniaki KAWABATA Hisato KOBAYASHI Pierre BLAZEVIC

RIKEN, Faculty of Eng., Institut des Sciences et
Saitama, Japan Hosea University, Techniques des Yvelines,
E-mail: kuniakik@riken.jp Tokyo, Japan Mantes, f i a n c e
E-mail: h@k.hosei.ac.jp blatevicQlisv.uvsq.fr

The purpose of our research is to develop the mobile system for rough terrain.
Our mobile system adopts the connected crawler mechanism. In this paper,
we proposed the operation strategy for autonomous motions. We also made
verification experiment of proposed operation strategy. For this verification, we
prepared 2 kinds of experiment. These experimental results showed that the
robot could pass over the different heights and different structures of obstacles
by using only (same) strategy. Moreover the sensors which realize proposed
strategy were very simple, and the number of sensor was very small. Therefore it
could be concluded that proposed strategies had an extremely high usefulness.

Keywords: Crawler, Rough terrain, Step climbing, Autonomous motion, Con-

nected crawler

1. Introduction
There is a great meaning to use crawler mechanisms as a mobile function on
rough terrain. Because, in general, the crawler mechanism can obtain big
impulsion due to its wider grounding bearing areas than the leg mechanism
and the wheel mechanisms. On the contrary, the crawler also has weak
points as a poor stability in complex geographical features. And the mobility
on the area like the stairs is inferior to that of the leg.'
Therefore, a lot of researches have been done to supplement these weak
points. The main theme common to those researches is t o improve the mo-
bility. Generally the method which changes the form of crawler is adopted

73
74

as an approach for this main theme. In order to realize this transformation,

many researches proposed the connected crawler mechanism which crawler
stages were connected by active joints. Lee et a12 designed two stages one
active joint type resucue robot. "Souryu-IIY3 is 3 stages 2 joints type for
resuce operations. "MOIRA"4 is also rescue robot which is 4 stages 3 joints
type connected crawler.
Although we can see such researches, there are no robots which show
autonomous operations. The one of the most important purpose to intro-
ducing rescue robots to disaster places is to automate sufferer searching
in place of the manpower searching. If many rescue robots can search the
sufferers automatically, it is enable to search wider and faster than conven-
tional manpower searching. However current rescue robots don't realize the
autonomous operations, therefore that is not achieved the above mentioned
important purpose which is to introduce robots to disaster places.
Thus this research proposes the rough terrain mobile robot which can re-
alize autonomous motion in disaster places. Especially, this paper proposes
operation strategies for passing over obstacles autonomously.

2. The prototype
First, let see the outline of our prototype. The mobile function of our pro-
totype adopts crawler mechanism. Because The crawler mechanism shows
the high mobile ability on various terrains; moreover it is simple mech-
anism and easy to control. But conventional single track mechanism has
mobility limitations; the limitation is determined by attacking angle, ra-
dius of sprockets, and length of crawler. In order to improve its mobility,
we add some active joints to conventional crawler tracks, namely that is
connected crawler mechanism. The specifications of the prototype are as
follows: Length(maximum) is 354.0 mm, Length(minimum) is 118.0 mm,
Width is 125 mm, Mass is 0.608 kg, Radius of the sprockets is 20.0 mm.

2.1. Mechanical s t r u c t u r e
Our mobile mechanism has 3 connected stages with the motor-driven
crawler tracks on each side(Fig. 1). RC-servo motors are used for driv-
ing joints between the stages. The left and right crawlers are driven by 2
DC motors independently. Three crawlers on each side are driven by same
motor simultaneously.
 75

Fig. 1. The overview of Connected

Fig. 2. The overview of the servo unit
crawler robot

2.2. Control structure

We adopt a hierarchical control structure by installing an intelligent servo
driver to each actuator. We connect each of them to the master contoll
unit by UART serial line. The parts marked by red line in Fig. 2 are servo
drivers.
Fig. 3 shows the control structure of this system. The master unit is
equipped with several sensors which are increnometers, PSD distance sen-
sor and photo reflector. The usage of these sensors will be shown in Chapter
3. Master unit culclates high level task (setting trajectory, sensing environ-
ment, etc), and servo driver works for low level tasks. The master unit
processes the high level task, and derive the data to low level task (crawler
velocity, joint angel and so on), and send them to the servo drivers. After
receiving these data, the servo drivers control their motor by conventional
feedback control low.

3. Operation strategies
The shpae of rough terrain such as disaster places is various. It is difficult
to derive the each autonomous motions relative to each terrain. Therefore,
in this chapter, we assume that the environment is one bump, and consider
about the operation strategies for autonomous motion to pass over the
bump. Because the climbing bump ability is important factor as one of the
most fundamental mobility index,5 climbing bump experiment is taken by
many researches as an evaluation experiment of mobility on rough terrain.
The proposed operation sterategy has 7 steps (Fig. 4). Following sections
show the details of each steps. Fig. 5 is the difinition of the parameters.
Here 8, is the orientation of the center stage. 81 and 82 are the 1st and 2nd
joint angle related to the 8,.
76

Servo unit

Fig. 3. The control system

Fig. 4. Proposed operation strategies

Rear Stage

.'...2M Joint

Fig. 5 . The definition of the parameters

3.1. First step

First, the robot goes forward until detecting the wall. If the robot faces the
wall, then robot stops moving. PSD distance sensor which is attached to
 77

the 1st stage for detecting the wall (Fig. 6). The output of the PSD sensor
is managed by the main controller (Fig 3).

,’ ,I

I* I*
** ,*

PSD distance sensor

Fig. 6. The PSD distance sensor Fig. 7. The definition of ffref

3.2. Second s t e p
In this step, 1st joint are driven to detect 8,,f. 8,, is the 1st joint angle
when the tangent of front stage meets the edge of the bump (Fig. 7).

3.3. Third s t e p
In third step, 2nd joint is driven while the robot goes forward. The purpose
of this step is to get the traction forces by keeping a grounding of rear stage.
If the robot goes forward without driving 2nd joint, then robot could not
get enough traction forces due to the lift of rear stage. In order to keep the
grounding of rear stage, 2nd joint angle should be set to angle of center
stage, namely the 2nd joint is driven in the following condition.

Here, the inclinometers are attached to the center stage for detecting the
angle of the center stage (Fig. 8).

Fig. 8. Inclinometers for detecting 6, Fig. 9. Contact detection device

78

3.4. Fourth step

In this step, 1st joint angle is set to 0 rad, 2nd joint is driven to let the angle
between rear stage and ground be right angle. At this moment, the robot
continues moving. There are two purpose in this step. One is to obtain the
traction forces, that is the role of 1st joint motion. The other is to lift up
the robot as high as possible, that is the purpose of 2nd joint motion. 2nd
joint angle is determined by following condition. By this condition, rear
stage can always stand with keeping right angle to the ground.

The trigger to shift third step to fourth step is the In the third step,
when the orientation of center stage Oc is equal to Or,,, operation step is
shifted.

3.5. Fifth step

The robot goes forward with keeping above mentioned conditions. When
the center of gravity of the robot is in the right side of the bump edge, then
the clock wise moment is generated around the edge, the robot can climb
a bump. Fig. 10 shows the situation of this case.

Fig. 10. Inclinometers for detecting Oc

3.6. Sixth and Seventh step

At the end, 2nd joint angle is set to the initial position, not to interfere
robot’s moving. The trigger for shifting to this motion is the contact be-
tween bump and rear stage. A photoreflector is attached to root and bottom
of the rear stage for detecting the contact(Fig. 9).

By above steps, climibing a bump is completed.

 79

4. Experiments
In order to confirm proposed operation strategy,We prepare two kinds of
experiment. One is that the robot passes over two bumps with different
heights. The other is stairs ascending. There are remarkable points in these
experiments. These experiments verify whether the robot can pass over the
different heights and different structures of obstacles by using only proposed
strategy. Moreover the sensors which realize proposed strategy is very sim-
ple, and the number of sensor is very small. Therefore if these experiments
success, it can be concluded that proposed strategy has extremely high
usefulness.

4.1. as sing over the diflerent hight bumps

In this section, the robot passes over the different heights of bumps. The
heights of bumps are 150 mm and 40 mm. We made the experiment by
implementing proposed strategy to main controller. The result is shown in
Fig. 11. This Figure shows that the robot can pass over the different hights
of bumps autonomously.

Fig. 11. The experimental results of passing over bumps

4.2. Stairs a s c e n ~ i n ~
Next experiment is stairs ascending. The height between stairs is 150 mm,
that is conventional stairs. The implemented software to main controller
is the same as experiment in 4.1, namely we do not add any modification,
that is completely same. The result is Fig. 12. From this Figure, it is turned
out that the robot could ascend stairs autonomously with driving joints.

5. Conclusions
The purpose of our research is to develop the rough terrain mobile robot
which can realize autonomous motion in disaster places. Especially, this
80

Fig. 12, The experimental results of stairs ascending

paper have proposed operation strategy for passing over obstacles au-
tonomously. This operation strategy is consisted of 7 steps, and it needs
only 3 simple sensors which were PSD distance sensor, inclinometer and
photoreflector. We have also made verification experiment of proposed o p
eration strategy. For this verification, we prepared 2 kinds of experiment.
One was that the robot passes over the different heights of bumps. The
other was stairs ascending.
Both experiments had a success. There were remarkable points in these
experiments. These experiments showed that the robot can pass over the
different heights and different structures of obstacles by using only (same)
strategy. Moreover the sensors which realize proposed strategy were very
simple, and the number of sensor was very small. Therefore it can be con-
cluded that proposed strategy has extremely high usefulness.

References
1. S. Hirose, Mechanical Designe of Mobile Robot for External E n ~ r o n m e n t s(in
Japanese), vo1.18, no.7, pp904-908 edn. (Journal of Robotics Society of Japan,
Tokyo, 2000).
2. C . Lee, Double -track mobile robot for hazardous enuironmen~a p p l i ~ t i o n s ,
vol. 17, no. 5, pp 447-495, 2003 edn. (Advanced Robotics, Tokyo, 2003).
3, T. Takayama, Development of connected crawler vehicle souryu-iii for rescue
application, in Proc. of 22nd conference of Robotics Society of Japan CD-
R O M , (Robotics Society of Japan, 2004).
4. K . Osuka, Development of mobile inspection robot for rescue activities:moira,
in Proceedings of the 2'009 IEEE/RSJ Intl. Conference on Intelligent Robots
and Systems, (IEEE, 2003).
5. T. Inoh, Mobility of the irregular terrain for resucue robots, in lUth Robotics
symposia, (pp 39-44, 2005).
CONTROLLING AN ACTIVELY ARTICULATED SUSPENSION
VEHICLE FOR MOBILITY IN ROUGH TERRAIN

SIDDHARTH SANAN
Mechanical Engineering, Indian Institute of science-Bangalore
Bangalore, Karnataka 560012, India

SARTAJ SINGH
Centrefor Artificial Intelligence and Robotics, DRDO-Bangalore
Bangalore, Karnataka 560093, India

K MADHAVA KRISHNA
Robotics Research Laboratov, HIT-Hyderabad
Hyderabad, Andhra Pradesh 500032, India

This paper studies the problem of traversing a rough terrain by wheeled vehicles. The
criterion for mobility of a wheeled vehicle in any terrain is formally developed, providing
insights into the mechanical structure requirements of the vehicle. A vehicle structure
with an actively articulated suspension is found as a solution to improved rough terrain
mobility. The contact forces of the vehicle with the surface being traversed are identified
as the critical factor in determining the traversability of the surface. Hence a control
strategy involving the control of the contact forces (normal and traction) is proposed. The
key feature of the locomotion strategy, thus, developed is that it provides a solution for
mobility in terrain which cannot be traversed using a solution involving assumptions that
ignore the dynamics of the main body of the vehicle.

1. Introduction
The problem of finding a suitable mechanical structure for a mobile platform
that can move efficiently on rough terrain not known a priori has been long
standing. The main disadvantage of the passively suspended and fixed
suspension vehicles is that the robustness of these systems under vaned terrain
conditions is not certain. To enhance the performance of such systems, a class of
robots with actively articulated suspensions called the Wheeled and Actively
Articulated Vehicles (WAAVs) has been developed, the terminology first used
by Srinivasan et al. [l]. The Hybrid Wheel-Legged Vehicles (HWLV) [2] is a
class of vehicles that consists of any combination of wheeled and legged
mechanisms. The Hylos [2] and WorkPartner [3] are examples of such vehicles.

81
82

A static analysis of the contact forces on an HWLV such as Hylos would

reveal that the contact forces are functions of body-weight, payload, posture and
contact angles [4], [5]. Hence previous work [6], [2] has aimed at controlling the
posture, defined by the internal configuration parameters, with the purpose of
optimizing the contact forces for maximizing traction and stability. Algorithms
for traction optimization and power efficient mobility in rough terrain are
presented in [4]. However [4] does not speak of conditions where forward
motion is not possible for a slow-moving fixed suspension vehicle.
The approach in this paper which we term as the force-control methodology
can be considered inverse of the approach as stated in previous work [2], [7]
with respect to traction for mobility in rough terrain. In that, we are able to
control the posture of the main body by controlling the contact forces rather
than vice versa. We aim at directly controlling contact forces that result in the
desired traction.
We develop an analysis of the surface conditions leading to the complete
loss of ability to move forward in a given terrain for any rigidly suspended
vehicle which provides the motivation for an actively articulated suspension
vehicle structure and also to develop our control methodology. The possible
advantages that such a system might offer as compared to the posture control
methodology based on the vehicle kinematics is also studied.

2. Related Works
Ch. Grand et al. at the Laboratoire de Robotique de Paris have developed the
Hylos robot [2]. The central theme for the Hylos locomotion is to achieve a
posture of the main body which maximizes stability and traction using the
posture control algorithm that uses the velocity model to set velocities at the
various joints based on the posture error.
The critical assumptions made in the analysis of the Hylos [2] are: (i)
sufficient contact forces exist to allow for pure rolling of the wheels, also
implying that wheel-ground contact exists at all times, and (ii) the dynamic
effects are neglected. The authors believe that the method of altering the contact
forces to control the posture and improve traction performance in an actively
articulated suspension would be a first attempt and hope to solve the problems
prevalent due to assumptions stated in [2].

3. Problem Definition
We need to find a method that controls the contact forcesF'at the various
contact points between the wheels and the surface such that the robot
 83

successllly traverses a terrain. Hence the primary motive of the vehicle is to

be able to traverse a given terrain. Since our aim is to study the traction or the
forward motion problem a planar analysis will suffice as also in [4],however
extensions to the 3D problem can be done. To start the development, we first
understand the need for WAAV as compared to a rigid suspension vehicle and
why it offers a better possibility to traverse a terrain. From this analysis, we
understand the role the contact forces play in helping the vehicle traverse a
given terrain. We then use a generic platform consisting of a main body and two
actuated wheels in the sagittal plane, each wheel connected to the main body
through a Linear Force Actuator for development of our control methodology.
We call this vehicle LFA-V (Figure 1).

Figure 1. The LFA-V mechanical structure

The following conditions are unsuitable for control methodology developed

in prior work [2]: (i) the vehicle is moving fast, (ii) large surface discontinuities
exist (surface does not belong to C'), and (iii) surface geometry that does not
permit mobility under static conditions of the main body of the vehicle. The
basic reason for the failure under these conditions is that any posture control
methodology based on the kinematic analysis of the system fails to sense its
interaction with the environment since all sensory data is local to the vehicle
itself. Interaction with the environment is necessary for maintaining near ideal
conditions such as pure rolling and no lateral slippage used for developing the
kinematic model to control the posture. This is one way in which the force
control methodology differs from the others.
In general, the normal force and traction are related by:
@>T>R+I (1)
where T is the tractive force, N is the normal contact force, R is the
resistance in the direction of motion and I is the inertia force. In most cases, I
is neglected. We therefore have to decide the required normal contact forces N j
at each wheel surface contact that limit the traction T, . It should be made clear
that resistance includes potential fields like gravity as well as non potential
84

fields like friction. We also need to device a control to achieve the desired
value of contact force FC = [ T NIT using an articulated suspension. This
issue has been addressed previously for industrial robots in a comprehensive
manner [8], [9] and the work described in this paper does not focus on this
issue. The control strategy developed in the paper is shown to be useful for
application to terrain conditions that are infeasible for traverse by fixed
suspension vehicles. For all further analysis, we assume the wheels have only
single point contact in the plane of analysis and that the wheels are cylindrical.

Figure 2. Wheel terrain contact forces and parameters

4. Fixed Suspension versus WAAV

Traction control systems are generally used in fixed suspension vehicles to
maximize the traction it gets from a surface, to prevent being rendered
motionless. The quasi-static equations that relate the normal contact and traction
forces to the forces on platform for a planar two-wheeled vehicle are given by
(2). As shown in Figure 2., the global fixed frame is { X Y Z } , where Y is
parallel to the gravity force; {xyz}, is the contact frame at the "i contact point
where T, is along x and N , along y ; V,' is the component of the vector from
the ithcontact point to the CG point in the] direction; and y, is the contact angle
at the ithcontact point. The index i increments as we move from the rear to the
front wheels.
A.C=D
For a vehicle with two wheels in the plane, A = A , .
cosyl -shyl cosy2 -siny2
shyl cosy1 siny2 cos y2
vi" -5" v; -v;
 85

C=[T, N , T, N , Y ; D=[F, Fy M,]T

These set of equations are redundant and hence infinite solutions exist
under the following constraints:
N , > 0 V i,i = {1,2} (3)

7"'"5 (T .r ) 5 7,' V i, i = { 42) (4)

T, 5 ,uN, V i,i = {1,2} (5)

rt"'=,r,min are the maximum and minimum torques that the motor at the "i
wheel can generate. For the vehicle to be able to move forward, it must at least
remain in equilibrium. Therefore, we set F, 2 0 , Fy =Wg and M , = 0 ,
neglecting effects like rolling friction, inertial forces. W is the mass of the
vehicle in kilograms and M , is the moment in the plane of analysis. The above
problem can be viewed as a linear programming problem, where a solution to all
variables exists if the solution space is not a null set. The objective function is
set as:

min(S), S = T, (6)

to evaluate the regions for y, and y 2 , where a solution to the above set of
equality and inequality constraints (2)-(5) does not exist. To demonstrate the
existence of regions with no solutions, we set y , = O and vary y 2 from 0 to
rr/2 . Also the pitch angle y is varied from 0 to z/3.
Figure 3 shows the plot of min(q) as a function of the contact angle y 2
and pitch angle y . The values of y2 and y corresponding to the brown
(darkened and flat) region indicates the region of infeasibility where no solution
for the given values of y z and y exists. In this region, no values of traction
forces at the wheels can maintain the system in equilibrium with the vehicle
having a rigid or fixed suspension. Hence if the vehicle is not moving fast, it
cannot overcome this state and cannot traverse the terrain. It is in these regions
of infeasibility that actively articulated vehicles such as the HWLV and the
LFA-V demonstrate their advantage.
A number of strategies have been identified with the actively articulated
class of vehicles to traverse regions that offer no solution with the rigid or fixed
suspension vehicles. In further sections we introduce the control methodology
devised and its use to implement these strategies.
86

Figure 3. Plot of min(T,) showing regions of infeasibility (dark flat surface)

5. Force Control Methodology

We use the planar model of the LFA-V in Fig. 1 described earlier to develop the
Force Control Methodology. The mechanical structure for analysis consists of 2
wheels, each pinned to an outer slide link which is connected to an inner slide
link through a prismatic joint. The inner slide is fixed to the main body. The
prismatic joint is actuated through a linear actuator, mounted on the main body,
to which a desired force F A can be commanded. This force F A acts between
the main body and the output slide. Although the input and output slides have a
finite mass, we consider this to be negligible compared to the mass of the body
and therefore neglect them in our analysis.
The 3 DOF’s of the system can be identified as the height h of the main
body and its pitch angle w and the position along the horizontal direction X,.
The wheel ground contacts are assumed to be no-slip contacts for the kinematic
analysis.
The initial aim of our control shall be to ensure that the vehicle is successful
in moving over a terrain for a particular posture ( h , ~ of ) the main body. The
position X, is defined by the motion of the vehicle. To ensure that the vehicle
moves in the forward direction, the force on the main body should be greater
than or equal to zero (F, 2 0) . This is the main criterion for the system to
traverse a given terrain.
We can set arbitrary requirements for the two parameters (h,v / ) assuming
they lead to solutions for and N , under constraints (3)-(5). In effect we try to
control the posture of the main body and its forward acceleration by controlling
the contact forces 7; and N , using the motor torques and the linear force
actuators.
We shall write down the quasi-static force balance equations for the LFA-
V. r’ represents the effective moment arm of the force at the ithcontact point
 87

acting perpendicular to t h e j direction. The fourth row in (7) has been included
to set equal traction forces at both the wheels which is a strategy used by most
automobiles. The equations below are written by considering the force actuator
force F,’ = N,/cos(Y, -ty) . These equations are different from that for a
rigidly suspended vehicle (2) shown earlier. C, D are defined previously in (2)

[;I.= (7)

For the LFA-V specific mechanical structure, A = A , .

In the following, c is a cosine and s is a sine hnction such that c3 = cosy/ ;
c,’ = cos(y, - ty) ; and other notations follow. Also

= kl sin(@,)cos(yl - ty) + r cos(y, - ty) ; 5”= k, cos( 0,)/cos( y, - y / )

v,’ = k, sin(@,)cos(y, - ty) + r cos(y, - y / ) ; v,” = - k, cos( @, )/cos( y 2 - ty)

0, = atan(21,/w); k, =

A, = c:.s3 c3 s:
-- ‘S3 c:.s, c3 s:
-- ‘S) ; B = [ l 0 -1 01
c: c:

Here r is the radius of the wheel, w is the wheel base of the vehicle and lI and
I, are the wheel-to-frame distance along the direction of the actuator of the first
and second wheel. We can control the posture and motion of the vehicle by
defining Fy , M , and F, .
Fy = k,& + kpeh -Iwg
- (8)
where eh = hd - h , and
- -
M , = hey+ kpey (9)

where e y = tyd - y
F, can be set to any arbitrary positive value for a desired acceleration.
As the next level of development to include constraints (3)-(5) an alternate
solution for {T,,N, ,Tz, Nz}, which is the solution of the minimization problem
88

min(S), S = 7', with the constraint equations A .C = D and inequalities (3)-(5)

is chosen. If a feasible solution space does not exist to the minimization
problem, then row 2 of (2) containing the equality constraint with Fy can be
dropped and the minimization problem can be set with the constraints:

A^.C=b (10)
and constraints (3)-(5). Use of constraint (10) is possible for the LFA-V since it
allows for dynamics at the main body in the vertical direction.
Figure 4 shows a plot of 7', as the solution of (6) with constraints (10) and
(3)-(5) using A , . It reveals that there is no infeasible set of conditions in the
scanned region for y , = 0 . Hence theoretically it is possible to traverse all
terrain by using the modified constraint equation (1 0). The solution would result
in dynamics of the main body in the vertical direction while ensuring the vehicle
moves forward.

Figure 4. Plot of min(T1) showing no regions of infeasibility

6 . Simulation Results
The simulation was done using Matlab Simulink and MSC Visual Nastran. The
surfaces used as terrain were higher-order cubic continuous surfaces such as B-
splines.
The main body maintained its height hd and pitch vd and the surface
variations showed little effect on the dynamics of the platform. Hence by the use
of force control, the dynamics/posture of the main body and therefore the entire
vehicle could be controlled. A minimum acceleration of 0.16 m/s2 was needed to
initially start the system which was commanded using F, in (5). This is due to
the faceted geometry of the rigid wheel used by the simulation package which
contributes in an analogous way to rolling friction in the physical world.
Maximum pitch angle deviation from zero was found to be -1.57" (Figure 5a).
 89

The rms of the pitch angle, w through the motion on the surface was found to
be 0.6727" as can be seen in Figure 5a. The performance was compared with a
rigid suspension. The maximum pitch angle deviation from zero was found to be
-5.73' in this case (Figure 5b) with rms value of 2.135'. Hence a large
reduction in pitch angle variation is possible by using force control in the LFA-
V. The spiked nature of the graph in Figure 5a is partly due to the faceted
geometry of the wheel mentioned earlier which causes oscillation of the wheel
centre. This is not visible in Figure 5b due to the scale of observation.

P i I RMS of Pitch Angle

1
time. f sec
a. LFA-V with Force Control
15
I
2
m - a
0
a
-8
-'!I 05 ' __Pitch

1'5
Af
RMS of Pdch Angle
2 2'5
time. f sec
b. Rigid Suspension
3 ' 35 !
Figure 5. Variation of pitch angle in the simulation

Figure 6. Time frames of the simulation with the LFA-V traversing an uneven surface'

The simulation time frames can be seen in Figure 6 . Hence the force control
methodology using the normal contact force and traction forces work
satisfactorily to maintain posture and ensure the vehicle traverses the terrain.
Thus an alternate solution to control the posture of a WAAV such as the LFA-V
is provided, based on the contact forces at the contact points, which can also be
utilized in terrain conditions difficult to traverse.

7. Conclusions and Future Work

It is formally understood that rigid or fixed suspension vehicles are unsuited for
successful operation under certain terrain conditions. This brings out the
primary motivation for using WAAVs to maximize traversability of the terrain.

* Simulation videos available at http://research.iiit.ac.in/-viswanath/robotics~orce-control.h~l.

90

Therefore a method based on controlling the contact forces responsible for

causing the motion of a WAAV like the LFA-V is developed. This provides a
strategy for successfully traversing a terrain and also controlling the posture of
the vehicle.
As future work, we plan to build a physical implementation of a system
utilizing the Force Control methodology. The force actuator element to generate
the required contact forces for the physical implementation has been identified
as a 2-dof leg with a wheel (leg-wheel) as the end effecter. Control for this leg-
wheel has been formulated to generate required normal contact forces, which
has been studied in past research. Also extension to deformable surfaces of
contact, particularly for outdoor field environments is needed.

References
1. S. V. Sreenivasan and K. J. Waldron, Displacement analysis of an actively
articulated wheeled vehicle configuration with extensions to motion
planning on uneven terrain, J Mech. Des. 118(2), 3 12-3 17 (1996).
2. Ch. Grand, F. BenAmar, F. Plumet and Ph. Bidaud, Stability and traction
optimization of a reconfigurable wheel-legged robot, Int. J. Robot. Res.
23(10-ll), 1041-1058 (2004).
3. A. Halme, I. Leppanen, S. Salmi and S. Ylonen, Hybrid locomotion of a
wheel-legged machine, in Proc. Int. Conf Climbing & Walking Robots
(CLAWAR), Madrid, Spain (2000).
4. K. Iagnemma and S. Dubowsky, Traction control of wheeled robotic
vehicles in rough terrain with application to planetary rovers, Int. J. Robot.
Res. 23(10-ll), 1029-1040 (2004).
5. M. H. Hung, D. E. Grin and K. J. Waldron, Force distribution equations for
general tree-structured robotic mechanisms with a mobile base, in Proc.
IEEE Int. Con$ Robot. Automat. (ICRA), Detroit, MI, USA, 4, 2711-2716
(1999).
6. K. Iagnemma, A. Rzepniewski, S. Dubowsky and P. Schenker, Control of
robotic vehicles with actively articulated suspensions in rough terrain,
Auton. Robot. 14(1), 5-16 (2003).
7. Ch. Grand, F. BenAmar, F. Plumet and Ph. Bidaud, Decoupled control of
posture and trajectory of the hybrid wheel-legged robot Hylos, in Proc.
IEEE Int. Conf Robot. Automat. (ICRA), New Orleans, LA, 5, 51 11-51 16
(2004).
8. J. T. Wen and S. Murphy, Stability analysis of position and force control
for robot arms, IEEE Trans. Automat. Contr. 36(3), 365-371 (1991).
9. L. L. Whitcomb, S. Arimoto, T. Naniwa and F. Ozaki, Adaptive model-
based hybrid control of geometrically constrained robot arms, IEEE Trans.
Robot. Automat. 13(1), 105-1 16 (1997).
 DESIGN AND CONSTRUCTION OF A ROPE CLIMBING
ROBOT

Juan Pablo Martinez Esponda'

Robotics, Automation and Educational Technology Research Group

Department of Mechatronics
Tecnoldgico de Monterrey Campus Santa Fe. Carlos Lazo 100. Col. Santa Fe,
Delegacidn Alvaro Obregdn C.P.01389, Mexico City, Mexico

We present details of a rope climbing robot developed as an undergraduate project, The

design was initially tested in a national level robotic competition but later modified and
improved significantly. The current working model is almost completely autonomous. It
can detect the rope automatically, closes its hands to grab the rope and can move up and
down using ultrasonic sensors to detect the arrival of ceiling and floor. All the power
sources are on-board and the system can sustain this load for a sufficient amount of time
to make it a valuable prototype for further work and improvements.

1. Introduction

Intelligent robots for the household or industrial repetitive tasks have been
under development for quite a long time with many commercial models now
available off the shelf. For example there are robots to clean the external
windows of sky rise buildings, robots to clean carpets in houses and offices, and
robots to carry load from one place to the other in manufacturing plants, etc. A
robot can replace many of the boring and tedious tasks disliked by human
operators. The cleaning of the external windows in large buildings is one such
area where considerable effort has been made to replace humans with robots.
The risk to the lives of human workers can also be eliminated provided we can
find robust automatic systems to replace humans in this type of work. In this
paper we present a robot developed specifically to climb up a rope and descend
automatically. The work was initiated in response to the Xh national mini-robot

..juanpas2k@yahoo.com.mx

91
92

competition held at Tecnologico de Monterrey - Campus Queretarot, in May

2005.
There were two categories in this competition in which one could
participate:

1. The participant could manually position the rope inside the robot
system and the robot was only supposed to move up and down the rope.
The rope was about 11 mm in diameter (meant for professional hiking
and climbing). In this case the maximum points were only 200, or

2. The participant could leave the robot at the extreme of a circle with
rope at the center. The robot in this case had to be autonomous to move
towards the rope, detect it, grab it, and finally ascend and descend it.
The targeted height was 2 meters. Such robots had 200 extra points
from the beginning and hence could go up to 400 points in total.

There were five participating groups in the year 2005 and the robot
designed by the author was the only one with the capability to move towards,
and detect the rope autonomously. However, the robot had some other problems
at the time of,the competition (which have now been resolved) and hence could
not ascend the rope. The problems were identified as follows:

1. Improper distribution of weight of the robot.

2. Badly designed relay based control system.
3. Heavy relays, contributing to the total weight.
4. Lack of experience, both in electronics and programming.
5 . Inability to cope with the rope thickness. It was not of a fixed diameter
and had variations between 1lmm to 16 mm.

The rethinking and redesigning needed to remove the above mentioned

problems and critical areas, took almost a year. The big advantage though was a
significantly increased knowledge of mechanics, analog electronics, digital
electronics, microcontrollers and their programming, control, etct.

+ www.qro.itesm.mx
* I t is important to mention that the author was starting his second semester in engineering when he
developed the first model and was still in the fourth semester when this paper was initially
submitted. Hence no formal training in the area of engineering design and robotics was available.
This is evident from the paper which is, for obvious reasons, scant in this area. The author will
start taking these courses starting kom next (spring 2007) semester.
 93

The plan of the paper is as follows. In section 2 we review the current literature.
Section 3 discusses our proposed design. We give our conclusions and
recommendations in section 4.

2. Review of current literature

This mini review contains two types of systems. First are the robots designed to
clean windows, walls, etc. These systems perform the same task as we
ultimately intend to do with our robot, though use many different strategies
rather than focusing on rope climbing. The second set of references discus
automatic climbers of different types. Since the goals are partly common in both
of these categories, we think such a review is justified, though not all the
techniques of either group serve our purpose.

The window cleaning robot WICTOR [I] uses a rail to move the system
across the window whereas the second subsystem cleans the window. The robot
remains hanging during this process. Hence part of the problem of sticking to
the window always without falling, is solved by hanging the robot thorough a
rope. The Fraunhofer Institute for Manufacturing Engineering and Automation
IPA in Stutgart, has designed various models for this purpose including
Raccoon. It can start from one comer of the window and can autonomously
move up and down and sideways to clean the whole window. It too uses suction
pads for this purpose. When idle in its base camp, it can also recharge its
batteries. Quirl [2] is a similar robot that uses suction pads to stick to the glass
window. It is a lighter version (only 600 grams compared to 6.5 kilograms of
original Raccoon) and has a footprint no larger than a postcard.

A lot of similar work in this area has been done by many other groups
around the world. For example Ref [3] discusses the ROBOCLIMBER spider
which can climb with weight on it. The applications of climbing robots in
maritime industries have been explored in Ref [4]. References [5-61 discuss the
effects of gravity and stability of climbing robot, respectively. Similar
applications, for example in the area of window cleaning and wall
climbing/inspection have been discussed in Ref [7] and Ref [8-91 respectively.

References [ 10-121 discuss the possibility of using a parallel robotic

manipulator as a climbing robot. The interest in this case is not in cleaning
windows but to develop a robot that can ascend complicated structures
94

describing unknown spatial trajectories, such as palm trunks, tubes, bridges, etc.
These papers suggest the use of a Stewart-Gough platform as a climbing robot
with the capability of avoiding structural nodes.

3. The New Modified Design

Based on the problems encountered we have made the following changes and
additions to the system:

Weight of the robot: The initial design was too heavy (with an
improper distribution of weight) and hence had difficulty in moving up
the rope. To rectify this, the chassis was perforated to the maximum
extent to reduce about 400 gms of weight. The circuits were redesigned
and built to reduce firther 300 gms of weight. To distribute the weight
correctly, some of the circuitry was relocated to make the system more
homogeneous.

Ultrasonic sensors: To make the robot more robust and autonomous,

ultrasonic sensors were added. The robot can now “see” the floor and
the roof and can also detect transparent surfaces.

Control system: Initially, the control system used very heavy relays and
needed a very high voltage to function properly. Hence this subsystem
was not optimized. The original controller was replaced by a Polulu5
system which is much smaller and controls the new motors. The
voltage requirement goes down to 12 volts. The new relays are much
smaller in size and need only 6 volts to operate and can transmit up to
30 volts at 10 amperes. The relays have been connected in series with
diodes so that they switch ON only when the voltage crosses a certain
level. The energy requirements go down to a level that can be managed
by normal battery cells.

The original motor has also been replaced by a better 18000 rev/min
motor which can generate sufficient torque with the use of a velocity
reducer. The final velocity goes down to 300 rev/min. This motor helps
the triangular part (Part 3 of Figure 1) to move up-

§ http://www.pololu.corn/
 95

5) The new circuitry uses a Basic Stamp" 2p40 and hence need a
regulator (Lm2940 in this case) so that it can work at the needed 5
volts. Problems with the serial communication between the micro and
the control circuitry were removed to make it work correctly.

After corresponding adjustments, the weight of the current model of the

robot is about 2.9 Kg including batteries. The angular velocity of the rotors is
about 0.73 rev/sec without load, The load, however, decreases the velocity
significantly to about 0.35 rev/sec.

Figures 1 and 2 show the back and front view of the robot and have been
explained in detail in the figures section of the article. One of the major parts of
the system, the pulley system which allows the system to grab the rope firmly in
place and helps it in moving up and down the rope is visible in figure 2.
Considerable amount design and construction time was needed to make the
system perfect for this type of application. The completed system has been
tested on many occasions including an on-campus demonstration.

4. Conclusions

In this paper we have presented and discussed the results of a project initiated
by a local robotic competition. It demonstrates a concept that can be extended
further and can be converted into a commercial product. The robot can only
ascend and descend at the moment and needs additional sub-systems for
cleaning purposes. With some modifications it can be used in various other
applications. The list includes fire rescue work in high rise buildings, vertical
mining, support in hikinghope climbing with extra weight, maintenance of
electric power lines, window cleaning and repair work, etc. Though the current
system is limited in its capabilities we hope to do a lot of extension work in the
current model in the future.

5. Acknowledgements

The author wishes to thank Tec de Monterrey - Santa Fe Campus for providing
the resources needed to complete this work.

References

http://www.parallax.com/
96

1. See, <http://www.iti.mu-luebeck.de/Research/NC/IES/CleaningRobots/>
2. See <www.fraunhofer.de/fhhg/Images/magazine1-2005-34f~tcm6-
14061.pdf>
3. R. Molfino, M. Armada, F. Cepolina, M. Zoppi, ”ROBOCLIMBER the 3
ton spider”, Industrial Robot. IF: Vol. 32 (2), 2005.
4. M. Armada, M. Prieto, T. Akinfiev, R. Fernandez, P. Gonzalez de Santos, E.
Garcia, H. Montes, S. Nabulsi, R. Ponticelli, J. Sama, J. Estremera, S. Ros,
J. Grieco, G. Fernandez, ”On the Design and Development of Climbing and
Walking Robots for the Maritime Industries”, Journal of Maritime Research
Vol. 1 (2) April 2005
5. T. Akinfiev (;) M. Armada, “The influence of Gravity on Trajectory
Planning for Climbing Robots with Non-rigid Legs”, Journal of Intelligent
and Robotic Systems, 35 (3), pp. 309-326. Kluwer, 2002. ISSN 0921-0296.
6. T. Akinfiev, M. Armada, M. Prieto, M. Uquillas, ”Concerning a Technique
for Increasing Stability -of Climbing Robots“, Journal of Intelligent and
Robotic Systems, vol 27,2000, pp. 195-209.
7. Houxiang Zhang; Jianwei Zhang; Guanghua Zong; Wei Wang; Rong Liu,
“Sky Cleaner 3: a real pneumatic climbing robot for glass-wall cleaning,”
Robotics & Automation Magazine, IEEE, Volume 13, Issue 1, March
2006 Page(s):32 - 41.
8. Longo, D.; Muscato, G., “The Alicidsup 3/ climbing robot: a three-module
robot for automatic wall inspection,” Robotics & Automation Magazine,
IEEE, Volume 13, Issue 1, March 2006 Page(s):42 - 50.
9. La1 Tummala, R.; Mukherjee, R.; Ning Xi; Aslam, D.; Dulimarta, H.;
Jizhong Xiao; Minor, M.; Dang, G., “Climbing the walls [robots],”
Robotics & Automation Magazine, IEEE, Volume 9, Issue 4, Dec. 2002
Page(s):lO- 19.
10. Roque Saltaren, Rafael Aracil, Oscar Reinoso and Maria Antonieta Scarano,
“Climbing Parallel Robot: A Computational and Experimental Study of its
Performance around Structural Nodes,” IEEE Transactions on Robotics,
VOL. 21, NO. 6, DECEMBER 2005.
11. M. Almonacid, R. J. Saltarh, R. Aracil, and 0. Reinoso, “Motion Planning
of a Climbing Parallel Robot,” IEEE Transactions on Robotics and
Automation, Vol. 19, No. 3, June 2003 485.
12. Aracil, R.; Saltaren, R.J.; Reinoso, O., “Climbing parallel robot: a robot to
climb along tubular and metallic structures,” Robotics & Automation
Magazine, IEEE, Volume 13, Issue 1, March 2006, Page(s):16 - 22.
 97

Figures

Figure1: The robot movers up and down the rope (1) using the motor (2). The
triangular part (3) moves up and down via the motor (6) and helps opening and
closing the hands (4) and ( 5 ) which have a strong spring attached on the upper
side to return to the original (closed) position. (7) is the part of the control
circuitry whereas the sensors to detect the ground are shown in (8). The back
wheel (9) is fixed in one direction though in principle it can be adjusted in any
desired direction. The Basicstamp is visible in (10).
98

Figure 2: This is the final form of the robot. The rope gets directed towards the
pulleys (2) using the system shown on the top (1). The sensors (3) detect the
arrival of the floor or some other obstacle while the robot is ascending. The
servos (4) control the folding hands (5) (shown opened here). These help the
rope stay in its position and start closing a few seconds after detecting the rope
so that the robot has moved further in the direction of the rope and the rope is
now in its position (1). This time has been measured and programmed according
to the servo speed. The sensor (6) shown as a small extended hand detects the
rope. The motor to open and close the rollers (2) is shown in (7), though it is
hardly visible. The electronic circuitry for the same is behind (8). The sensors (9)
detect the arrival of the floor or any other obstacle. The orange roller (10) helps
the system move forward and backward. The other four rollers are there to
facilitate the movement and maintain tension in the belt. This belt helps the
system avoid slipping.
DESIGN OF NEW LEG MECHANISM FOR A WHEELED WALL
CLIMBING ROBOT *

YI-LI FU,HE-JIN YANG

Mechanical Engineering School, Robotics Institute, Harbin Institute of Technology
Harbin, 150001, China

High mobility and strong surface adaptability are two basic requirements for a counter-
terrorism wall climbing robot. However, traditional wall climbing robots fails to meet
both of the two requirements. This paper designed a new leg mechanism and by fixing it
onto the front of a wheeled wall climbing robot, obtained a novel wall climbing robot
which can move quickly on both smooth and rough planar surfaces, make transitions
between them and stride obstacles when necessay. Various aspects about the leg
mechanism’s design are introduced, including the design requirements, mechanical
design and gait design .Simulation and experiment results demonstrate the validity of the
design concept.

1. Introduction
In recent years, there have been strong demands for counter-terrorism wall
climbing robots to perform reconnaissance and surveillance missions as
terroristic activities increase. Because we never know when and where the crime
is going to happen but have to respond to it quickly, we need the robot should
be of high mobility and strong surface adaptability, which offers big challenges
to the locomotion mechanism design. Traditional wall climbing robots, no
matter what applications they were designed for, can be divided into two major
categories according to their locomotion mechanism: continuous motion robots
[ 1--31 and alternate motion robots [ 1,4--71. The continuous motion robots have
high mobility but poor surfaces adaptability because they usually move on
wheels. On the contrary, the alternate ones can be adapted to irregular surfaces
but very slow, for the simple reason that they use their legs or other adhesion
mechanisms in turn. Therefore, traditional wall climbing robots are not
completely satisfactory for counter-terrorism application.
This paper proposes a new concept wall climbing robot different from
traditional ones: It combined a wheeled single suction cup wall climbing robot
with an innovatively designed leg mechanism, created a robot that can move

* This work is supported by the National Natural Science Foundation of China (60675051).

99
100

both continuously and alternately, i.e. the robot uses its wheels on planar wall
surfaces with a fast speed, when comes across on obstacle, the leg mechanism
will be used to stride it. We envision that this robot be used in urban
environments for surveillance purpose. This proposal is reasonable because
most of the building surfaces are flat except for some obstacles such as gaps,
window frames andor steps. The main job of this paper is to introduce the
design of the leg mechanism.
This paper is organized as follows: Section 2 gives general requirements to
the leg mechanism design. In section 3, detailed descriptions about the
mechanical design are introduced. Section 4 illustrates how the robot strides
obstacles and makes surface transitions. In section 5, we present the experiment
and simulation results. Conclusions with future works are drawn in section 6 .

2. Design Requirements
As mentioned in section 1, mobility and surface adaptability is two essential
requirements to meet. Our strategy is to employ an existed wheeled robot and
design a leg mechanism for it, so the mobility will not be a problem.
Considering the wheeled robot’s payload and that the new robot will be self-
contained, controlled wirelessly, with half-autonomous intelligence, the
requirement for the leg mechanism design are:

Be as an independent module
Be able to work on both smooth and rough surfaces
Strong striding obstacle ability
Can make transitions between walls
Light
Simple structure
Energy saving
Easy to control

3. Mechanical Design
Great efforts have been made to meet all the design requirements. Figure 1
shows the final model of the new robot. It consists of two parts: the body and
the leg mechanism. The body is actually a typical wheeled single suction cup
robot as we mentioned earlier. The locomotion and adhesion mechanisms are
similar to the robot in Ref.[2].
The leg mechanism is designed depending on the size and weight of the
body listed in Table 1. As an independent module, it did little changes to the
 101

body. The leg mechanism is made up of 3 links and 3 joints. In robotics theory,
it is a 3-DOF planar manipulators thus is very easy to control. The specification
of the leg mechanism is also listed in Table 1.

Figure]. A CAD model of the new robot with the leg mechanism

Table 1.specificationsof the body and the leg mechanism

I Size I3 0 0 ~ * 3 0 0 ~ * 8I 5 ~

The suction cup located at the end of the manipulator, which is part of link
3, is special designed for rough surfaces. As shown in Fig.2, when the suction
cup is covered on a rough wall, two types of apertures can be shaped: open
aperture and closed aperture. Because of the suction cup’s large size and wide
edge, most apertures belong to the closed ones. However, some open apertures
102

still exist. It is the flexible rubber strip that crams them so that a vacuum
chamber can be formed. If the apertures are too large or too deep to be crammed,
they could be treated as gaps and will be bypassed or stridden.
Rubber Sealing Strip

on Cup

.--?
0 .,
. .:z:. -.>

a: open apertures b closed apertures

Figure2. Illustration of the suction cup’s sealing mechanism

4. Gait Design
“Gait” refers to how a legged robot moves its legs. It is very important to
investigate the gait because it helps to determine the leg’s structure and size,
select joints’ driving motors and to h o w the obstacle dimensions that this robot
can deal with. The gait was designed following two principles:
1. Safe: the mass center of the robot should be close to the wall so that the
robot would not fall down under the gravity;
2. Less power consumption: less power c o n s ~ p t i o nmeans longer working
time and lighter batteries, which is very important to a self-contained wall
climbing robot.

a b C

Figure3. The process of striding a obstacle and making transitions

Figure 3(a) illustrates a typical process of striding a step or a gap. If the
distance between the small suction cup and obstacle is longer than the length of
body, the process could be simplified as shown in Fig.3 @).
 103

Figure 3(a) illustrates a way of making wall to wall transitions. It is similar

to the process of striding obstacles but the joints’ dnving torque is much smaller.
This process could also be regarded as ground-wall, wall-roof, wall-ceiling
transitions. But the wall-ceiling transition is not realized because the body is too
heavy.
The prismatic joint in the leg mechanism offers the robot powerfbl abilities
to stride obstacles. Figure 4 illustrates several different situations and the
admissible conditions. The numbers are based on the robot size in Table 1.

a b C d e
e 2 393 e 2 150 ( e 2 286)&(h 584) t (e2 450)&(h < 306)
Unit: mm
(150 < e < 450)&(h < 306)
“&” means “and”

Figure4. Illustration of striding obstacles and admissible conditions

5. Simulations and Experiments

Both kinematic and dynamic simulations were made to assist the
mechanism’s design and to verify the design concept. The simulation model we
built is much identical with the real robot. Here we present one of those results
as shown in Fig.5, which records the distance between each link’s mass center
and the wall surface during the process illustrated in Fig.3 (a).The result
indicates the succeed of the planned path and the satisfaction to the safety rule
in Section 3.
Some simple experiments were made to test the adhesion ability of the
suction cup designed in Section 3.We used the micro-pump in Table 1 to
produce vacuum and got a maximum negative pressure of -38kpa on glass wall,
and -32kpa on brick wall. According to the size of the suction cup, a minimum
of 2595N force and a 194 Nm moment could be produced in theory. In fact, a
15kg object was hung at joint 3 and the suction cup never slipped or fell down.
The results indicate that the suction cup can adhere to both smooth and rough
walls reliably.
104

Center-of-Mass
375 0
350 0
325 0
300 0
275 0
2500
1E 2250
s 200 0
1750
a" 1500
125 0
100 0
15 0
50 0
25 0
00 30 60 90 12 0
mme (sec)
Figure5 The trajectories of each link's mass center

6. Conclusions
A new concept robot which is a combination of the wheeled wall climbing robot
and the successfully designed leg mechanism is proposed in this paper. It
overcomes the limitations of previous ones and surpasses them in terms of
mobility and surf-ace adaptability. The leg mechanism succeeds in many ways
such as strong surface adaptability, simple structure, modularity and so on. The
future work is to design an embedded controller board and integrate the
electrical system with mechanical design to realize the surveillance purpose for
counter-terrorism missions.

References
1. A. Nishi, Development of wall-climbing robots. Computers Elect. Eng.
22(2): 123-149 (1996).
2. Y. Wang, S.L. Liu, et a/, Development & Application of Wall-Climbing
Robots. Proc.IEEE Int. Conf Robotics & Automation. (1999).
3. J.Z.Xiao, A. Sadegh, et al. Design of Mobile Robots with Wall Climbing
Capability. Proc.IEEE/ASME Int. Conf Adv. Intel. Mechatronics (2005).
4. B. Luk, A. Collie and J. Billingsley, ROBUG 11: An intelligent wall
climbing robot, Proc. Int. Con$ Robotics and Automation, (199 1)
5. A. Nagakubo and S. Hirose, Walking and running ofthe quadruped wall-
climbing robot. Proc. IEEE Int. ConJ robotics and Automation. (1994).
6. H. R. Choi and S. M. Ryew, et al. A Wall Climbing Robot with Closed Link
Mechanism.Proc. IEEE/RXJ,Int.Conflntelligent Robots and Systems. (2001)
7 . R. L. Tummala, et al. Climbing the Walls. IEEE Robotics and Automation
Magazine. pp 10-19, Vol. 9,No. 4 (2002).
DEVELOPMENT OF A CLIMBING ROBOT FOR INSPECTION
OF LONG WELD LINES

JIANZHONG SHANG, BRYAN BRIDGE, TARIQ P. SATTAR, SHYAMAL

MONDAL, ALINA BRENNER

Centrefor Automated and Robotic NDT, London South Bank University,

I03 Borough Road, London, SEI O M , UK

This paper presents a wheeled robot designed for climbing on ferrous surfaces for the
inspection of real time inspection of long weld lines simultaneously with the welding
process. Neodymium permanent magnets are used for adhesion, which are capable of
producing a maximum adhesion pressure of 4x104Nm” at a 20mm air gap. The strong
neodymium magnets make the robot have a high payload canying capability. The
arrangement of the magnet array increases its performance at large air gaps so that the
robot has excellent capability to overcome obstacles, such as weld caps. The design of
the wheeled robot with two sections jointed by a hinge joint takes the advantages of high
speed and good manoeuvrability, as well as working on curved surfaces and transferring
between angled surfaces.

1. Introduction

1.1. Objectives of the project

The project Climbing Robot Cell for Fast and Flexible Manufacture of Large
Scale Structures (CROCELLS) seeks to modernise and take into the future the
technology of the manufacture of large fixed welded structures such as box
girder bridges, storage tanks, ships and other steel fabrications which arise on
construction sites, in large chemical and foodstuff plants, on offshore oil
platforms etc. It creates a transportable manufacturing cell consisting of a team
of cooperating climbing robot work tools whose activities are coordinated and
integrated through a central intelligence. Each robot will be dedicated to a
different task to optimise overall system performance[ 11.
The system in this project consists of three cooperated robots, a welding
robot holding a welding torch and its control unit, a utility robot that carries the
weld wire spool and navigational sensors and an NDT robot for carrying weld
inspection tools. As part of the project, the NDT robot for weld inspection is

105
106

developed in the Centre for Automated and Robotic NDT, London South Bank
University.
The main purpose of the NDT robot is the real time inspection of long weld
lines with 100% volume coverage, simultaneously with the welding process.
Such NDT data is then be used to control the welding process parameters
including authorisation of weld repairs in real time. Some examples of the real
work environment, assembly of Blocks and/or Mega Blocks to form the ship,
ship hull of container ships, are shown in Figure I.

(b)
Figure 1 Examples of work environments

1.2. ~ e n e r u~l e q u i r e m e nof

t ~the ~D~ robot
The requirements of the robot mainly include:
* The robot should be able to climb on steel walls up to 30m height.
Climbing performance on curved surfaces where the ~ n i m u mdiameter is
down to 2.5m.
* System weight for an individual robot platform, including the robot weight
and payload, is required to be less than 70kg.
 107

To surmount surface obstacles, such as ridges on ship hulls, of up to 20mm

height. Higher obstacles, for example stiffeners and nozzles will be
circumnavigated.
To transfer between angled surfaces (135 degree) such as those which arise
on container ships.
To move between stiffeners of 700mm separation such as those on container
ships, therefore the width limitation of the robot with NDT probes mounted
is 600mm.
The robot is required to move at constant speed to follow the welding
process as well as move at high speed when locating itself to different places
to save time. The welding speed for ship hulls and storage tanks are about
10 to 15m per pass per hour.

1.3. Current Technology and the Challenge

There are a number of ways for the adhesion of a wall-climbing robot, such as
vacuum suction cups, magnets, sticky materials, etc. For a steel-wall-climbing
robot, the obvious adhesion technology is magnetic force. A great number of
robots with magnet adhesion have been developed by researchers in different
organisations. They are commonly featured as robots with magnet tracks[2,3,4],
robots with magnetic wheels/disks[5] or legged robots with magnet feet[6,71, etc.
However, none of them can meet the requirements of the particular application
in this project. A new robot has been designed and the real challenge of the robot
development is the capability that the magnets work over large air gaps for the
purpose of overcoming obstacles and working on curved surfaces.

2. Design of the NDT Robot

In order to achieve the smooth and continuous movement, as well as excellent
manoeuvrability of the robot, a wheeled robot with magnet adhesion is
developed.

2.1. Magnets investigation

The payload of the NDT robot is approximately 10kg. To surmount the obstacle
of 20mm height, the minimum air gap between the magnets and the surface is
20mm, which requires bigger magnets to generate the sufficient holding force
over the air gap. However, big magnets will increase the robot weight. In
consequence, the payload carrying capability is decreased. The trade-off of the
size, weight and magnetic force are considered carefully.
108

A comparison of a 1 by 4 magnet array with different gap between the

magnets is illustrated in Figure 2. The attraction force is between the magnet
array and a big steel plate with lOmm thickness. The result shows that when the
magnets are closer to each other, more attraction force is created at further
distance, and the force variation from tested distance (5mm to 35mm) is less,
which is our desired result.
From Figure 2, we also see that the use of a back plate significantly
increases the attraction force, especially at further distance. The use of mild steel
back plate has the following advantages:
Flux concentration. With the steel back plate, more magnetic lines are
concentrated between the back plate and the steel wall, so that the attraction
force is increased.
Reduce the magnetic field strength in the back, so that there is less
interference to the electronic equipments carried on the robot, which makes
the shielding of the magnetic field easy.
Ease the management of the magnets. The adhesion force can fix the
magnets on the steel plate so that there is no additional fixing mechanism
required.

1600

1400

1200

g 1000

:
-
800
-2Ommapanwilhbackpble
-5mmapanwilhaul bach piale

c 600

400

200

0
10 20 30 40
Air gap (mm)

Figure 2 Comparison of attraction force at different magnets arrangement

Based on the analysis above, two blocks of magnet array are built as shown
in Figure 3. Instead of using one single big block magnet, which is extremely
difficult to handle, a number of smaller block magnets are placed on a mild steel
plate to form a magnet array that covers big area. Each of the magnets is made of
neodymium 42 and has the dimension of 50mm x 50mm with 25mm thickness.
 109

In each magnet array, the magnets are placed against each other as close as
possible, so that more magnetic lines are forced to the working surface.

Figure 3 Magnet arrays with neodymium block magnets on mild steel plates

The magnetic force of the designed magnet array against the increasing of
the air gap is shown in Figure 4.

Magnetic force vs air gap

5000

4500
4000

~
- 3500

1 3000 _.
-3 x 3 anv
~ 2500
-r-3
__ x 2 anw

I 4 1500

too0

500

I lo
Air gap (mm)
2o 30 40

Figure 4 Attraction force changing against air gap

2.2. Robot Structure

The NDT robot is designed to have two sections connected by a hinge joint, with
two wheels to drive the robot, and two omniwheels, one in the front and one in
the back, to support the robot. The driving wheels are driven by two servo
motors differentially, so that the robot can be turned easily and driven along any
direction. The structure of the robot is shown in Figure 5.
110

Magnet mavs
Figure 5 Robot structure

The two-section design is concerned for the robot travelling through the
sharp angled corners presented in ship hulls. In the design without the hinged
structure as shown in Figure 6(a), when the robot transfer from one plane to
another at the corner, the air gap increases too much so that the magnet force is
too low to support the robot.
With the hinged design as shown in Figure 6(b), when the front half of
magnet is lifted up, the back half still remains strong holding force. When the
driving wheels reach the comer, the front magnets resume strong holding force,
then the back magnet is lifted up to complete the transfer.

Figure 6 Illustration of the hinged structure that can transfer between angled surfaces
 111

The driving wheels are made of aluminium hubs bonded with solid 65 IRHD
polyurethane tyres. The polyurethane material has the coefficient of friction of
0.9 on steel walls. The key benefits of the material are
Abrasion Resistance (Non-Marking)
Impact Resistance
Cut Resistance
Heat and cold resistance
LoadBearing
Long Working Life
High Grip-Good Drive Material
Oxygen & Ozone Resistant
The robot is featured as follows:
Size: 600mm x 375mm x 340mm
Weight: 30kg (Including payload)
Speed: 0-lOm/min
Minimum gap between magnets and work surface (flat): 20mm

2.3. Robot Control

The NDT robot control unit is developed as illustrated in Figure 7.

I Wireless adaptor I
Central Management
Computer I Power supply

Figure 7 Robot control unit

112

The on-board robot controllers include two motor control modules and one
interface conversion module. All the modules have Ethernet interface, so that
they can be controlled via Ethernet network. The interface conversion module
converts Ethernet to serial, 10, AD and 12C interfaces, allows the connection of
different sensors and equipments, enables monitor or control them via standard
TCP-IP protocol, makes the combination with other equipments easy.
All the on-board modules with the standard Ethernet protocols are plugged to
an Ethernet hub which is carried on the robot. The uplink of the hub is connected
to the wireless bridge for the wireless communication with the central task
manage computer. The wireless communication is based on the Wi-Fi
technology. The application of Wi-Fi wireless communication eliminates the
communication cable, so that the robot umbilical is reduced. The only umbilical
for the robot is the power supply cable, which cannot be eliminated at this stage
due to the robot is required to work for long time.

2.4. Robot Guidance and cooperation with the welding robot

For the robot travelling along the stiffeners such as those which arise on
container ships, there are two infrared distance sensors facing side way, one in
the front and one in the back of the robot. The sensors measure the parallelness
and distance between the robot and the stiffener. According to the sensor
feedback, the controller calculates the speed of the two driving wheels to keep
the robot moving parallel to the stiffener and maintain a constant distance.
In the case of hot weld inspection, the robot needs to follow the welding
robot. A thermal array sensor, as in Figure 8 is used together with the infrared
distance sensors to follow the hot welding point. The thermal array sensor has an
array of eight thermopiles arranged in a row. It can measure the temperature of 8
adjacent points simultaneously. The sensor reads infra-red in the 2um to 22um
range, which is the radiant heat wavelength.

Figure 8 Thermal array sensor for guiding the robot

 113

To follow the hot welding point, the NDT robot is placed next to the welding
robot as shown in Figure 9, with a reflective object raised on the welding robot
so that the NDT robot can 'see' the welding robot. The infrared distance sensors
read the distance from the welding robot and the central controller keeps the
NDT robot parallel to the welding robot and remains constant distance. The
thermal array sensor is placed with its row of the thermopiles parallel to the
weld. By controlling the robot so that the hottest point always stays at one
particular sensor element, the robot will always follow the hot point where the
welding torch points to.

(C)
Figure 9 The NDT robot follows the welding robot for hot weld inspection
114

3. Conclusion
The NDT robot has been developed with the capability of climbing on steel
walls carrying the specified payload and the capability of overcome the
obstacles. The robot is also able to climb on curved surfaces with excellent
manoeuvrability, and transfer between angled surfaces.

Acknowledgments
This work is sponsored by the European Community through the project
Climbing Robot Cell for Fast and Flexible Manufacture of Large Scale
Structures (CROCELLS), 6th Framework Programme, project No.
STRPO17509[1] with the following partnership: London South Bank
University(UK), Cybernetix(France), Zenon(Greece), Bisiach and C&(Italy),
NAlTOSOL(Greece), TERNA and GEK Group(Greece). The project is
coordinated by Professor Bryan Bridge of London South Bank University.

References
1. Deliverable D2 report for the European Community project Climbing Robot
Cell for Fast and Flexible Manufacture of Large Scale Structures
(CROCELLS), 6th Framework Programme, project No. STRP017509.
(Restricted access to the partnership and the EC only).
2. W. Shen, J. Gu and Y. Shen, Proposed Wall Climbing Robot with
Permanent Magnetic Tracks for Inspecting Oil Tanks, Proceedings of the
IEEE, International Conference on Mechatronics & Automation (2005).
3. Z. Xu, and P. Ma, A wall-climbing robot for labelling scale of oil tank's
volume, Robotica, ISSN:0263-5747, Pages: 209 - 212 (2002).
4. L. P. Kalra and J. Gu, An autonomous self contained wall climbing robot
for non-destructive inspection of above-ground storage tanks, Industrial
Robot: An International Journal, Vol. 34, No. 2, pp. 122-127 (2007).
5. S. HIROSE, H. TSUTSUMITAKE, Disk Rover: A Wall-Climbing Robot
Using Permanent Magnet Disks, Proceedings of the IEEE/RSJ International
Conference on Intelligent Robots and Systems (1992).
6. B. L. Luk, A, A. Collie and J. Billingsley, ROBUGII: An intelligent wall
climbing robot, Proceedings of the International Conferance on Robotics
and Automation, pp. 2342-7 (1991).
7. T. Kang, H. Kim, T. Son and H. Choi, Design of quadruped walking and
climbing robot, IEEE/RSJ International Conference on Intelligent Robots
and Systems, Vol. 1, pp. 619-24 (2003).
 DEVELOPMENT OF A SEALING SYSTEM FOR A
CLIMBING ROBOT WITH NEGATIVE PRESSURE
ADHESION

C. HILLENBRAND', D. SCHMIDT and K. BERNS

Robotics Research Lab,
Department of Computer Science,
University of Kaiserslautern,
67653 Kaiserslautern, Germany
*E-mail: cahillen@informatik.uni-kl.de
http://agrosy.informatik. uni-kl.de

T. LEICHNER, T. GASTAUER and B. SAUER

Institute of Mechanical Engineering and Gear Technology,
Department of Mechanical and Process Engineering,
University of Kaiserslautern,
67653 Kaiserslautern, Germany
http://megt. mv. uni-kl. de

The non-destructive inspection of large concrete walls is still an unsolved prob-

lem. One possible technique is to use driven wheels for the propulsion and a
vacuum system for the adhesion. The seals for the vacuum chambers are s l i p
ping over the rough surface, therefore it is not guaranteed that the chambers
are always airproof. Especially over concrete walls a special seal construction
must be found to make the adhesion more safe. On the other side the propul-
sion system must be able to produce enough force for carrying and accelerating
the robot to a suitable velocity.
This paper will present the climbing robot CROMSCI which uses the de-
scribed techniques. The propulsion system consists of three omni directional
driven wheels which are airproof and completely rotatable and has been pre-
sented in earlier papers before. For adhesion a vacuum system of seven con-
trollable vacuum chambers and one reservoir chamber is used. This system
including chambers and seals will be discussed in more detail. The rough and
sharpedged surface of concrete walls cause strong requirements to the sealing
concerning leak tightness and attrition. Therefore, each sealing must be flexible
to allow a good adaption to the ground but also let the robot slip when the
wheels are turning.

Keywords: Climbing Robot, Adhesion System, Negative Pressure, Sealings.

115
116

1. Introduction
Regular inspections and repairings of concrete buildings like motorway
bridges and dams are very extensive. Among high expenses in money and
time the technical stuff has to work in hazardous environment by using
complex access devices. For a better objectivity and reproducibility of in-
spections as well as for safe working conditions a climbing robot has been
build up which can check the building remote controlled. During different
experiments with the prototypes the importance of sealings became obvi-
ous. Although we got promising results onto smooth concrete or wooden
surfaces the sealing failed on rough surfaces which can be found on every
concrete building. For this the existing adhesion concept was overhauled
and new sealing materials and constructions were developed.
In literature other climbing robots using negative pressure adhesion can
be found, although not all of them can be used for this application [l][2]
[3] [4]. Some of them are suitable only for flat surfaces like glass or use
legs for locomotion, which will result in slow movement. For climbing on
concrete surfaces an active vacuum system driven by wheels seems to be the
best solution because of fast continuous motion and a simple mechanical
structure.

2. Adhesion Concept
The adhesion concept of our climbing robot CROMSCI” consists of seven
single vacuum chambers which are supported by one large reservoir cham-
ber a t the top of the robot. Each chamber receives its negative pressure
from the reservoir chamber which is evacuated by three suction engines
and can be controlled by valves. If one or more working chambers are loos-
ing negative pressure they can be isolated from the vacuum system to avoid
the propagation of normal pressure to the other chambers which will re-
sult in loss of adhesion. The air-pressure in each chamber and the reservoir
is measured by pressure sensors. These informations are given to a pres-
sure controller which opens and closes the valves t o evacuate the chambers
separately depending on the actual leak tightness. The components of the
adhesion system are:

0 one large reservoir chamber (including valve to outside)

0 three suction engines evacuating the reservoir chamber

&ClimbingRobot with Multiple Sucking Chambers for Inspection tasks

 117

e seven working chambers (each including valve, pressure sensor and

sealing)
e electronics for low-pressure control
Beside the real adhesion mechanism a thermodynamical model has been
created to simulate the airflow and pressure variation with modelled leakage
areas and cracks. It shows that the usage of multiple sucking chambers are
as important as a good driving strategy if the cracks are too large for the
sealing. More details on the vacuum control system can be found in 151.

2.1. New Sealing Design

An important aspect is the concept of the sealings between the negative
pressure inside the vacuum chambers and the normal pressure outside. To
guarantee that the climbing-robot keeps the contact to the wall while mov-
ing on the surface, it is neccessary to develop a seal which needs to be
wear-resistant, leak-proof and easy sliding. To reach this intention a pro-
totype demonstrator has been developed which enables to apply several
different sealing materials and shapes to find the optimum in friction and
wear.

valve tox, txofiles

Fig. 1. Prototype and drawing of the basic construction

It is an assembly of one seal-bearer and two aluminium profiles, in which

the bearer is clipped (fig. 1).It is possible to vary the height of the seal-
bearer by changing the air pressure in the tube made of Butyl. Onto the
surface of the tube the different samples of seal material and friction can
be applied.
118

Another important point with regard to practical application is an easy

changing of the seals if they become porous due to abrasion. With the
shown sealing this replacement can be done fast and simple.

2.2. Test Stand

To detect the lift/drag ratio and the value of friction between the sam-
ple and the surface vis-a-vis, a Pin-on-Disc test stand is used which fulfills
DIN 53516 [6]. With this test bench it is possible to adjust different ve-
locities of the turning disk to simulate the relative movement between seal
and wall. The normal force, which is affected because of the vacuum to pull
the robot against the wall, can be simulated with the help of some weights,
which are placed on top of the pin. The sliding friction is detected by using
a beam in bending, which is coevally the reference surface for the inductive
sensor to detect the amplitude of the beam. To determine the wear across
the time of testing, a second inductive sensor is applied on the opposite site
of the beam in bending. They are both placed on one rocker, shown in Fig-
ure 2. The sliding friction and the wear will be detected while using a dry,
wet and steamed contact between pin and disk. In this way, all conditions,
which appear on the seal of the climbing robot, can be simulated.

inductive sensor (friction) pin and disc

inductive sensor (wear)

Fig. 2. Pin on Disc test stand

3. Locomotion
As already mentioned before our robot CROMSCI is driven by three omnidi-
rectional wheels which let it slip over the surface. The interaction between
adhesion system and drive is important due to the fact that the robot should
 119

neither stuck to the wall nor fall down. If the sealings are too soft the robot
would suck itself to the wall, if they are too hard they lose leak tightness.
So we have to make a crompromise and allow controlled leakage.
To measure all occuring forces a load cell is integrated into each wheel
so that the robot can recognize if the negative pressure is too high for
movement or dangerously low. The drive and the adhesion system will
soon be integrated into one control mechanism which allows a safe driv-
ing onto rough surfaces. More informations about the propulsion system
can be found in [7].

4. Experiments and Test Results

In the first studies with the Pin on Disc stand the friction between several
samples of seals are tested. In the following chart, the different parameters
can be seen.

No.
Test I1 Velocity
rel-
El
Temp-
erature
[“CI
Lubri-
cation

A
Contact
Pressure
~ o - ~ N
Material
Pin
Material
Disc

0.1-0.03 20 I - 9.4 Rubber Glass

2 I 0.1-0.03 20 Rubber Marble
3 I 0,l-0.03 20 Rubber Concrete
20 - I 9,4 Carpet Glass
20 Carpet Marble
20 Carpet Concrete

The tests no. 1 to 3 are made with rubber on glass, plane marble and
plane concrete. They provide a basis to compare well known materials with
the samples of carpet. In all experiments the temperature and the contact
pressure were kept constant. It could be possible, that the temperature
departs in a field of f l kelvin but this is irrelevant for the result of the
friction coefficient. It was attempted to run the test without any kind of
lubrication, but this was not possible with the samples of concrete. The
concrete disc wears very fast, because the pin in contact grinds the fine
sand on the surface. This fine sand works like a kind of lubrication and
tempers the dry friction. However all pins, which stay in contact with the
concrete disc, excite this special wear, so they are comparable among each
other. Three different materials were used in the Pins: rubber, carpet (short
fibres) and vertofloor (long fibres) as shown in figure 3.
120

Fig. 3. Disc materials (top) and pin materials (bottom) used for the tests

o.'2 --T-

0.02

o 10
/ - I - - --x
50 70 m

Fig. 4. Dynamic friction: rubber to glass

Figure 4 shows the graphs of the velocity steps and the value of the
friction coefficient between rubber and glass. The dynamic friction coeffi-
cient starts with a value of approx. 1 at the highest relative velocity of 0 , l
meters per second and rises slowly to 1," while the velocity decreases to
0,03 meters per second. The irregularities in the graph are generated by
disturbances in the contact, resulting by the symmetry of the disc.
Figure 5 shows similar to figure 4 a friction coefficient. In opposition the
current graph does not follow a clear line, so the dynamic friction cannot be
clearly seen. Because of the distinctive swinging of the results, it has to be
assumed, that there is more static than dynamic friction. That's why the
trend line is not in the middle of the amplitudes, but on the maximums.
 121

0' b -1

10 30 50 70 90 110 130 150

Time [S]

Fig. 5. Static friction: rubber to plane marble

0.12 , IS

01
10 30 50 70
_--
90
-.
110 130
1 2
150
Time [s]
'-v[mlpl-?i;b$;'

Fig. 6 . Static friction: rubber to concrete

The contact between rubber and concrete (fig. 6) seems like a mixture
of figure 4 and 5. The results of the experiments are also swinging, but
the dynamic and frequency is much lower than in figure 5. Similar to the
friction coefficient in the graph before, the dynamic friction coefficient can-
not clearly be detected. Again the static friction has to be used, to get a
comparable result.
The graphs in figure 7 are the first results, which can be compared to
the graph in figure 4.Two distinctive differences can be seen: all values are
much lower. They vary from 10% to nearly 30%. The second conspicuous-
ness in the graph is the inverse pitch: the rubber friction rises with falling
velocity, the carpet frictions are nearly constant and fall a t the end of the
experiences. The carpet slides much easier on glass than rubber, whereas
122

110 SeJ 1

Fig. 7. Dynamic friction: carpet t o glass

the carpet with the short fibers has a lower friction coefficient.

I_

Fig. 8. Dynamic friction: carpet to plane marble

By comparison figure 5 with figure 8 the friction and the swing frequency
between adhesion and sliding is much lower. With the graph in the current
figure it is possible to find a dynamic friction, which is significantly lower
than the static friction of the rubber in contact with the plane marble. The
difference is again nearly up to 10
The course of the graphs in figure 9 shows a distinctly static friction.
It is necessary to use the maximums of the amplitudes to compare it with
figure 6 . The course of the friction for both kinds of carpets is approx. from
a third up to a quarter. Summary of results:
 123

12

I
d
.
oa =
B
0,6
0
0.4 s'
0,2 B
0

001 i
I F I
. -02
0 I -04
10 30 50 70 go 110 150 150
Time [s]
1 -
Fig. 9. Static friction: carpet to concrete

Test No. Static Friction Dynamic Friction

1 rubber 1,l - 1,7
2 rubber 2,3 - 2,7
3 rubber 3,7 - 4,9
4 carpet 0,33 - 0,37
4 vertofloor 0,20 - 0,25
5 carpet 0,17 - 0,25
5 vertofloor 0.32 - 0.37
6 carpet I 0,70 - 1,00 I
6 vertofloor I 0,60 - 1,22 I

5 . Conclusion
In this paper we introduced a novel sealing system which meets high de-
mands. A sealing prototype is presented and different applicable facings
have been tested concerning friction and abrasion onto several surfaces like
concrete or glass. These experiments pointed out, that the sliding character-
istic strongly depends on the facing material and will have a high influence
on robot movement and adhesion.
Future Work mainly consists of experiments under real conditions. For
this the adhesion system of our robot CROMSCI (see figure 10) has to be
completed so that it will cling to the wall. Furthermore more experiments
concerning the leak tightness of the sealing facings have to be carried out.
These results will be compared with simulated experiments and evaluated
to achieve best adhesion attributes.
124

Fig. 10. CROMSCI platform as presented at the Hannover Messe 2007 with seven vacuum
chambers, mounted sealings (left) and reservoir chamber (right).

References
1. Y . D. Zhang, K. L. Fan, B. L. Luk, Y . H. Fung and S. K. Tso, Mechanical
design of a climbing cleaning robot, in 8th IEEE Conference on Mechat~nics
and Machine Vision in Practice, (Hong Kong, 2001).
2. D. Longo and G. Muscato, Design of a single sliding suction cup robot for
inspection of non porous vertical wall, in ISR 2004: 35th Inte~ationalSym-
posium on Robotics. Proceedings., (Paris, France, 2004).
3. M. Hagele, Assistive robots in everyday’s environments Europe’s Information
Society - ICT Riga 2006(June, 2006).
4. B. L. Luk, D. Cooke, S. Galt, A. A. Collie and S. Chen, Robotics and Au-
tonomous Systems 5 3 , pp. 142(0ctober 2005).
5. J. Wettach, C. Hillenbrand and K. Berns, Thermodynamical modelling and
control of an adhesion system for a climbing robot, in 20th IEEE Inte~ational
Conference on Robotics and Automation (ICRA), (Barcelona, Spain, 2005).
6. Oenorm din 53516: Testing of rubber and elastomers - determination of abra-
sion (1987).
7. C. Hillenbrand and K. Berns, The force controlled propulsion and adhesion
system for a climbing robot, in 9th Inte~ationaZConference on Climbing and
Walking Robots (CLAWAR), September 12-14 2006.
 Development of a Suction Type Miniature Climbing Robot with
Minimal Actuators

M.V.Vignesh
Research Scholar, Department OfMechanical Engineering, Anna University,
Chennai-25,India.

L.Karunamoorthy
Professor, Department OfMechanical Engineering, Anna University,
Chennai-25,India.

ABSTRACT

A wall-climbing robot intended for inspection of tall structures alike has been developed. The
robot has characteristic features of kinematical design and is capable of moving a tool at a specific speed
to follow complex paths on surfaces. In real field conditions, the labor intensive inspection demands great
attention to details by inspection personnel and is subjected to human errors and limited in reliability. The
primary objective in the design of this vertical crawling robot is focused on its ability to travel vertically
on a wall. The robot totally uses two actuators, one for actuating the trunk and the legs to get the forward
motion, backward motions and lifting up and down the suction cups. The other one is to give the rotary
motions for changing the directions. Thus the robot uses totally two actuators for navigating on the curved
surface as well as along the surface of the vertical wall. In this work, it has been successfully tried to
actively employ the body as an extra leg.

1. Introduction

Wall- climbing robots are modeled to do their jobs in place of physical labor.
Miniature climbing robot developed for inspection in harsh conditions can reduce the
labor-intensive work and man-hours. Such robots should be small enough for easy
handling and capable of traveling on vertical, inclined and curved surfaces of any
structure. Hence, an attempt has been made to develop a novel multifunctional
automated crawling system, which can carry the miniature instrumentation to
perform a wide range of inspection tasks while attached to the surface of the
structure of interest.

A large size robot of heavy weight with permanent magnetic disk wheels, which is
capable of steady and smooth omni- directional locomotion on the surface of a flat or

125
 126

curved iron wall, has been developed by Shigeo et a1 [ 11. A robot, whose motions are
accomplished by means of flexible pneumatic cylinders made of rubber, has been
developed by Leoncio et a1 [2]. A robot of larger size with a heavy weight capable of
moving a tool at a specific speed to follow a complex path is developed by Lin Guo
et a1 [3]. A multifunctional automated crawling system to carry miniature
instrumentation to perform a wide variety of tasks while being attached to an
aircraft’s surface is developed by Paul et a1 [4]. A self contained robot developed by
Tomoaki et a1 [ 5 ] uses suction for gripping. A robot with a magnetic suction
arrangement, which can work in extremely hazardous environments, but with the
limits of many factors has been developed by Wang Yan et a1 [6]. A robot which
uses a caterpillar concept to perform 3D complex movements to inspect a high
number of dangerous manual operations especially in large industrial environments
has been developed by Abderrahim et a1 [7]. Ming Chen and Song Huat Yeo [S]
have developed a robot planar walker, based on a novel planar 8-bar mechanism
using planar rigid motion group. Two climbing robots considering under-actuated
kinematic structures have been developed by Mark A. Minor and Ranjan Mukherjee
[ 9 ] . A quadruped robot designed by Taehun Kang et a1 [lo] has used an integrated
approach to carry an ultrasonic tool for inspection of large surfaces in industrial
utilities. Jizhong Xiao et a1 [ 111 have developed a gait controlled miniature bipedal
climbing robot with under actuated mechanism. A quadruped robot that uses wheel
legs for locomotion on surfaces regardless of the presence of gravity or direction was
developed by Daltorio et a1 [12]. But most of the previous wall-climbing robots have
utilized a number of actuators which are generally composed of complicated bulky
mechanisms for driving and adhesion which lead to the increase in volume and
weight accordingly.

Therefore, an attempt has been successfully made to design a miniature-

climbing robot with minimum number of actuators. The robot thus developed uses
only two actuators. The work presented in this paper describes only a part of the
ongoing program on developing miniature robot, with critical mechanism to make
the robot climb vertically with minimum number of actuators deployed in it.

2. Mechanical Structure

The mechanical arrangement of the robot shown as an exploded view in Fig.la

consists of three main assemblies, namely a frame, a drive mechanism and two legs.
 127

The photographic view of the robot is shown in Fig.lb. The robot is 165x188~90
mm in size made of aluminum weighing 1500 g capable of carrying a pay load of
800 g. The crawler employs stepper motor for mobility and suction cups for surface
adhesion. The crawler has legs for linear motion and a rotary element for curvilinear
motion.

The main frame, namely the trunk is fitted with two actuators one for actuating the
trunk and the legs to get the forward and backward motions of the robot and the
other one to give the rotary motions for changing directions. Thus the robot uses
totaIIy two actuators for navigating on vertical wail. The trunk and the legs have
been fitted with separate suction pads on them. Suction pads help in surface adhesion
for robot. Robot uses totally seven suction cups, three for the trunk and the
remaining four for the legs, having two suction cups on each leg. The location and
the placement of the suction cups are shown in fig.2. Vacuum generator activates the
suction cups, which uses compressed air for functioning. In this work, it has been
tried to actively employ the body as an extra leg. This idea becomes more effective
because it can produce a simple gait pattern. The suction pads of the trunk are
activated and fixed to the wall. By rotating the actuator, the legs move in a
semicircular trajectory and close to the wall as described in fig3 and at the same
time the vacuum sensors detect whether the vacuum at the Legs is estabIished or not.

1.Leg one, 2.Motor one, 3.Motor two, 4.h;mk,5. Rotary element, 6 . Leg two
Figure la. Exploded view of the robot
 128

Figure lb. Photographic view of the actual robot

Once the legs are firmly Iixea to tne wail aue to me vacuum createa in xnem, the
suction pads of the trunk are released. The ongoing rotation of the motor makes the
trunk to step forward to have a forward motion, as shown in Fig.3 and now again the
suction pads in the trunk will have a contact with the wall. Forward motion of the
robot is produced by successively repeating this sequence of operations. The same
actuator which is used for forward and backward motion of the robot is used for
lifting up and down of the suction cups. This lifting up and down of the suction cups
is very much necessary to avoid dragging of them

As the robot is developed to perform specific tasks like inspection, the mechanical
structure includes even special fittings on it and suitable NDT sensors intended to
inspect the structuxal flaws can be fitted on to the robot body. These fittings are
custom built for different types of sensors depending on their size and application.

Figure 2. Bottom view of the robot showing the location of the suction cups
 129

Stage Stage Stage3 Stane Stage

* *
.r

Suction ‘3 Release

Figure 3. Gait control for linear motion

3. Gait Control

The sequences of translation and rotary motions are shown in Figs.3 and 4. To obtain
a forward motion, the trunk is initially made to stick to the surface (Fig.3). Gait
control for rotary motion then the two legs are simultaneously stepped forward and
are made to fN to the surface. Once the legs are fixed to the surface the trunk
releases from the surface and it moves forward. The forward motion of the robot is
obtained by successively repeating this sequence of operations. To make the robot
steer on the wall, the trunk is made to attach to the surface. While the two legs are
detached from the ground, they are in a raised position. Now an actuator is used to
rotate the trunk about a point, which enables the robot to turn around and make its
move. The Fig.4 shows the stages and the positions of the suction pads at the time
the rotary motion is taking place. Thus the robot body will negotiate a directional
change using the steering actuator, when the suction pads of the trunk are attached to
and the legs detached from the wall.

Stage Stage 2 Stage Stage 4

a b F

a
a. 5Ma one b. side two F. Robot front
Suction Release

Figure 4. Gait control for rotary motion

 130

4. Walking Algorithm to Scan an Area

A climbing robot can be used for inspection of surfaces. The quality of the results
obtained from the scanning of an area, depends upon the resolution of motion of the
robot. A 1 mm x 1 mm scanning resolution is considered to be optimal .To obtain
this, the robot motion should be made more precise and accurate. By moving the
trunk and the legs, the robot is capable of obtaining a forward motion as shown in
fig.3. A spring loaded probe holder is used to obtain constant contact with the test
surface. This makes it possible to get 1.0 mm resolution along one direction. To get
the same 1.0 mm resolution along the other orthogonal direction, the robot is made
to follow a set of movements as shown in Fig.5. After keeping the trunk of the robot
firmly made to stick to the wall, firstly the robot legs are lifted from the ground.
Secondly the robot is made to rotate through an angle of 2.388' followed by two
steps. The forward motion of the trunk covers a linear distance of 24mm. The
resultant of these two motions gives the required 1.0 mm linear distance in the
orthogonal direction. By reversing the same procedure, which is by rotating the robot
through 2.388' in the reverse direction, the robot is brought to the original straight
position. Thus similar operations are repeated to get the linear movements in the co-
ordinate directions.

Starre 1 Stage2 Starre Stage 4

Figure 5 . Scanning algorithm for orthogonal movement

5. Pneumatic Control

The Fig.6 shows the pneumatic circuit developed for wall-climbing of the robot.
This pneumatic circuit is used only for gripping the surface. The pneumatic circuit is
separated into two parts; the suction cups and the vacuum efficiency valve are placed
on the robot while the circuit components are placed in the ground station. The
 131

separation of circuit components is shown in fig.7. These separated circuit parts are
connected by flexible tubes.

Compressed air is let into the system by opening a shut off valve. The air from the
compressor is filtered before it is made to use, thus let air pressure is regulated and
maintained using regulator. Here the air channel is split into two routs one will take
care of the suction cups in the legs and other will take care of suction cups in the
trunk. In each channel, electrically operated solenoid valves will regulate the air
flow. These solenoid valves are controlled by pc based software. Vacuum in the
suction cups is achieved, depending upon the opening or closing of the solenoid
valves. Once the solenoid valve is opened the vacuum generator is made to work.
Suction is achieved through the vacuum generator based on the Venturi effect.
Compressed air is forced into the vacuum generator and it is made to pass through
and out of the exhaust. On its way through, it draws air in from a perpendicular pipe
causing suction at this pipe. This results in gripping of the suction pad. Vacuum
efficiency valve are put to use, when any of the suction cups fail to have grip on the
surface. That is, in some situations a suction cup may not be able to stick to the
surface. Under these situations, the robot motion will be stopped. To avoid such
situation, vacuum efficiency valve will shut the mouth of the suction cup to prevent
any more air leakage. Even if any one of the suction cup fail to stick to the surface,
with the help of other suction cups the robot can be put in motion. The airflow is
regulated through solenoid-actuated valves. Depending upon the position of legs and
trunk, the valves act accordingly. The controller guides the stepper motor and the
solenoids in a sequential manner to achieve the vertical motion of the robot.

6. System Control

The overall system layout sketch is shown in fig.7. Labview based software is
used for controlling the robot. The software controls both the stepper motor and the
pneumatic circuit. The robot can be semi automatically made to reach the scanning
location from where the robot can be fully automated by software to scan the desired
surface. In case of any failure in achieving the surface grip, the software can be
intimated through vacuum sensors so that the robot does not move any further. The
solenoid valves and the vacuum sensor are controlled through homemade hardware
which is too software controlled.
 132

8 I

il ~ - -Leg
-- _ _ _ _
Trunk
._
_ I
!
1. Shut off valve 2.Filter regulator 3.Solenoid valve 4. Vacuum generator
5 . Vacuum efficiency valve 6 . Vacuum Sensor 7.Suction cups
Figure 6 . Pneumatic circuit

On robot

-4-

Figure7. Overall system layout

7. Conclusion

In this work, a wall-climbing robot is designed and made with compact and
robust mechanism. Compared with the previous wall climbing robots that are
available in literature, the present robot has the unique advantage of having reduced
number of actuators and other parts as well as with an easy controlling module. It is
developed in such a way that the robot can produce all the elementary motions like
the forward, backward and the rotational with use of only two actuators. The same
actuator which is used for forward and backward motions of the robot is used for
lifting up and down of the suction cups. The special mechanical structure allows this
robot to explore environment using climbing gaits. The pneumatics is used only for
 133

gripping the surface, while the stepper motors are used for mobility in this robot. The
dynamics of the robot is electrically controlled through a computer using a code
developed in Labview for this purpose. Such a crawler can travel on an aircraft
structure and automatically perform inspection tasks that can be controlled remotely.
Effectively, this robot can be used to integrate a suite of sensors and to perform
intelligent scanning and testing for corrosion and other defects. The crawler is made
at significantly more affordable cost by employing commercially available
components while allowing the NDE industry to concentrate on making dedicated
sensor modules as plug-and-play boards. The crawler is developed to operate
autonomously and search for flaws on structures without human interference.

The components of the robot namely, main body, worm gear drives, crank
drives, legs, hub drive and relevant smaller parts like bushes have been fabricated.
The stepper motors and all other components like compressor, suction cups, solenoid
valves, vacuum generator required for pneumatic system have been purchased from
the market.. The software for the drive has been developed using Labview to suit the
requirement. The assembly of the robot has been carried out using all these
components and successfully demonstrated for its smooth mobility in crawling with
a sufficient pay load.

8. Reference

1. Shigeo Hirose, Hiroshi Tsutsumitake and Disc rover: “A wall-climbing robot

using permanent magnet disks”, Proceedings of IEEE, International Conference on
Robotics and Automation, Raleigh, NC, July 1992. pp. 2074 - 2080.

2. Leoncio Briones, Paul Bustamante and Miguel A. Serna, “Wall-climbing robot for
inspection in nuclear power plant”, Proceedings of IEEE, International Conference
on Robotics and Automation, May 1994, Vol. 2, pp. 1409-1414.

3. Lin Guo, Kevin Rorers and Robin Kirkham, “A climbing robot with continuous
motion”, Proceedings of IEEE, International Conference on Robotics and
Automation, Santiago, May 1994, Vol. 3, PP. 2495-2500.

4. Paul G. Backes, Yoseph Bar-Cohen and Benjamin Joffe, “The multifunction

automated crawling system (MACS)”, Proceedings of IEEE, International
 134

Conference on Robotics and Automation, New Mexico, Apri1997, Vol. 1, pp. 335-
340.

5. Tomoaki Yano, Shinji numao and Yukio Kitamura, “Development of self-

contained wall climbing robot with scanning type suction cup”, IEEE, International
Conference on Robotics and Automation, Canada, Oct 1998, pp. 244 - 259.

6. Wang Yan, Shao Hoa, Lliu Shuliang, Xu Dianguo and Gao Xueshan,
“Development and application of wall-climbing robots”, IEEE, International
Conference on Robotics and Automation, Michigan, May 1999, V01.2, pp. 1207 -
1212.

7. M.Abderrahim, C.Balaguer, A Gimenez, J.M.Pasto1 and V.M.Padron, “ROMA: A

climbing robot for inspection operations”, Proceedings of IEEE. International
Conference on Robotics and Automation, May 1999, Vol. 3, pp. 2303 - 2308.

8. 1.Ming Chen and Song Huat Yeo, “Locomotion and navigation of a planar walker
based on binary actuation”, Proceedings of IEEE. International Conference on
Robotics and Automation, Washington, DC, May 2002, pp. 329-334.

9. Mark A. Minor and Ranjan Mukherjee, “Under-actuated kinematic structures for

miniature climbing robot”, ASME Journal of Mechanical Design, June 2003, Vol.
125, pp. 281 - 291.

10. Taehun Kang, Hyungseok Kim, Taeyoung Son and Hyoukryeol Choi, “Design of
quadruped walking and climbing robot”, Proceedings of IEEE. International
Conference on Robotics and Automation, Las Vegas, Oct 2003. pp. 619 - 625.

11. Jizhong Xiao, Ning Xi, Jun Xiao and Jindong Tan, “Multi-sensor referenced gait
control of a miniature climbing robot”, Proceeding of IEEE, International
Conference on Robotics and Automation, Las Vegas, Oct 2003, PP. 3656-3661

12. Kathryn A.Daltorio, Andrew D.Horchler and Stanislav Gorb, Roy E.Ritzmann
and Roger D.Quinn, “A small wall-walking robot with compliant, adhesive feet”,
IEEE, International Conference on Robotics and Automation, July 2005, pp. 34 - 39.
 DEVELOPMENT OF AN AMPHIBIOUS HEXAPOD ROBOT
BASED ON A WATER STRIDER

SOH FUJI1
Major of Precision Engineering, Graduate School of Science and Engineering,
Chuo University,
1-13-27 Kasuga, Bunkyo-ku, Tokyo 112-8551, Japan

TARO NAKAMURA
Major of Precision Engineering, Graduate School of Science and Engineering,
Chuo University,
1-13-27 Kasuga, Bunkyo-ku, Tokyo 112-8551, Japan

There is an increasing demand for mobile robots that can operate in a variety of
environments, including across the ground, on water, and over rough terrain. In this study,
we focused on a water strider, which is an amphibious creature. The water strider moves
on water, supporting its body on four legs and propelling its body with two other legs. On
land, water striders walk using all six legs. We developed an amphibious hexapod robot
based on the water strider, and confirmed that this robot can move on the ground and on
water. This robot can move smoothly over shoals because it does not use a screw
propeller.
The locomotion mechanism of the water strider robot has 12 servomotors. It is equipped
with a radio camera and controller, which allows operation from a remote location.
Furthermore, this robot can change its direction while remaining in place.
Experimental evaluation of the performance of the water strider robot, moving on both
ground and water, showed it to be similar to that of an actual water strider.

1. Introduction
Today’s mobile robots are expected to function in a variety of environments,
including land, water, and irregular terrain. Mobile robots, both bipedal walking
robots and those with wheel drive, are coming into practical use; however, their
movement mechanisms limit their use in more difficult environments. There is
still a demand for robots that can move freely to disaster locations and places
that are difficult or dangerous for humans to enter. For instance, if a robot can
freely move on both water and land, it would be useful for rescue operations and
research in desert locations.
In this study, we examined a water strider, which is an aquatic insect that is
able to move easily on water and on land. In addition, it can move smoothly in

135
136

shallow and shoal areas, and make sharp turns. We developed a water strider
robot, which is a mobile robot with the locomotion characteristics of the water
strider, with the goal of realizing an amphibious movement mechanism that
affords both high mobility and stability [11[21.

2. Locomotion of a living water strider

.A water strider has fuzzy legs covered with wax that makes them water
resistant. The high surface tension of these legs makes the water strider float on
the surface of the water. The water strider is small, and weighs from O.Olto 0.lg.
The slim leg creates a wide depression on the surface of the water, isolating the
water strider’s body from the surface of the water. Figure 2.1 shows the
depressions the legs create on the surface of the water.

Fig. 2.1. A water strider on water.

A water strider puts its fore and hind legs on the water to support itself and
uses the two mid legs for propulsion. These mid legs scratch the water like the
oars of a boat. Changing the angle of the legs controls the direction of
movement and speed. In addition to supporting the body, the hind legs are used
as a rudder, providing minute steering action. The front legs, in addition to
providing support, can grasp an object for predation.

3. Experimental plan
We evaluated the locomotion mechanism of the water strider and developed
an amphibious robot. The robot has six legs, and each leg has the same
specifications; however, the tips of the front and hind legs are equipped with
floats for floatation and the mid legs have oars. We simplified the water strider’s
movement mechanism. During water movement, the front and hind legs are
locked in place, and the mid legs are used as the drive legs-moving the mid
legs drives the robot. On land, however, the robot raises the mid legs, fixes them
in place, and uses the front and hind legs as its drive legs. Moving the front and
hind legs provides the mobility. Independent servomotors drive each leg. A PIC
controls the servomotor’s angle of rotation. In addition, the robot has a power
 137

supply, built-in control system, and radio control. Figure 3.1 shows the robot,
and its specifications are given in Table 3.1.

Table 3.1. Amphibious robot

specifications.

Fig. 3.1. An amphibious robot.

We examined mechanisms for lifting the legs and for moving the legs back
and forth. The mechanism that moves a leg from front to back consists of the
turning of a servomotor. Figure 3.2 shows a schematic of the leg mechanism.
The mechanism for lifting the leg was realized with a parallel linkwork, such as
@ in Fig. 3.2.
Servo motor
Forward to backward

-43
0
Servo motor

Fig. 3.2. Operating the robot leg.

Servomotor @ lifts (or lowers) the leg, and servomotor @ moves the leg
from front to back. In other words, servomotor @ allows the leg to rake the
water.
138

4. Locomotion in water

4.1. Gait model

In water, the robot fixes its two front and two hind legs, and moves only on the
mid legs. The same method is used by the water strider. Here, we implemented
four movements: forward, backward, tum to the left, and turn to the right. These
correspond to the front, rear, left, and right commands on radio controller sticks.
Figure 4.1 shows the movement of each leg as the robot moves on water.

Advancement FJ
I_I:sustain legs F]
:moving leg

Fig. 4.1. Advancement on water.

As the robot moves on the water, it moves its mid legs from fiont to back.
The front and hind legs do not move, but support the body.
Once the robot moves ahead, it moves the mid legs in reverse motion to that
shown in Fig. 4.1.
Next, I examined movement in shallow areas. Figure 4.2 shows the
movement of each leg when turning left on the water.

Turning to left j lcyclc 1

I_I:sustain legs 0 :moving leg

Fig. 4.2 Turning left on water.

To turn left on the water, the left mid leg moves from back to front and the
right mid leg moves from front to back, as shown in Fig. 4.2. The fiont and hind
legs do not move, and only support the body.
To turn to the right, the robot moves its mid legs in the opposite direction.
 139

4.2. Experimental apparatus

We performed experiments to evaluate the robot’s motion on the water.
Figure 4.3 shows the experimental setup. I fixed a camera with a tripod and
photographed the robot’s water movements from directly above. We used
MOVIAS pro3D for analysis,just as we did for the water strider.
We analyzed the movement of my new amphibious robot, and compared its
performance with that of an actual water strider.

Camera

I / \ I

Fig. 4.3. Experimental apparatus for evaluating locomotion on water.

4.3. Experimental results

Figure 4.4 shows the movement distance and angle of the amphibious robot’s
mid legs.

I -Left oar -Right oar -Center of the body I

40 I 1 0.3
30 0.25
-2 20
10
0.2 -E
zw 0.15
2 o
-10 0.1 g
-20 0.05
-30 0

0 0.5 1 1.5 2
time [sec]
Fig. 4.4 Amphibious robot performance data
140

This Figure shows that mid legs move forward and backward by 30 degrees. For
this progress, the right and left mid legs move in a motion that is linearly
symmetric to the body axis. The body retreats slightly in reaction to the forward
motion of the mid legs, but once the mid legs row, the robot moves forward.

-Right oar -Left oar Center of the body

0 20 40 60 80 100
time [msecl
Fig. 4.5 Water strider data.

Figure 4.5 shows the movement of an actual water strider.

The robot moves 30 mm with one stroke of its mid-leg oars. The robot body
is 410-mm long and the body of the living water strider is 14-mm long. The
increased power of the robot is no larger than the increase in resistance from
upsizing the volume, and the differences in its floating method produce the
surface tension (“the ascending force”).

5. Locomotion on the ground

5.1. Guit model

To move on the ground, the robot uses two front and two hind legs. The mid
legs are raised so that they do not touch the ground and are locked in place
[31[41[51. The robot fixes three of the four remaining legs, and moves by operating
the remaining leg, selecting the leg to move, based on the direction of travel.
This is called a “crawl gait,” and is a basic gait of quadrupeds. The robot has
four kinds of advancing movements: front and back legs is fixed, turning to the
left, and turning to the right.
Figure 5.1 shows the leg movements for movement on land.
 141

I I

Advancement I lcyclc I
:sustain legs 0 :moving leg

Fig. 5.1. Leg movement on ground when advancement

When walking on four legs and turning to the left, the mid legs are fixed
(above the ground) and three of the front and hind legs provide balance while
the last leg provides propulsion. The moving leg changes in rotation.
Figure 5.1 shows that the moving leg changes from left front to right hind,
right front, and finally left hind. These represent one cycle.
Figure 5.2 shows leg movement for turning left on the ground.

Turning to left j lcyclc I

I_I:sustain legs 0 :moving leg

Fig. 5.2. Leg movement on ground when turning left.

When walking on four legs and turning to the left, the mid legs are fixed
(above the ground) and three of the front and hind legs provide balance while
the last leg provides propulsion. The moving leg changes in rotation.
Figure 5.2 shows that the moving leg changes from left hind to right hind,
right front, and finally left front to complete one cycle.
Turning to the right involves the reverse rotation of these motions.
142

5.2. Experimental apparatus

We performed an experiment to evaluate the robot’s mobility on land. We
fixed a camera with a tripod and photographed the motion of the robot on land
from directly above. The analysis method is similar to that in water.

5.3. Experimental results

Figure 5.3 shows the position transition of each leg and the body’s center of
gravity as the robot moves.

I I I I
0
0 0.2 0.4 0.6 0.8
%xis hil
Fig.5.3. Analysis of movement on land.

Each leg lifts, moves rorwarci, and then mags bacltward slowly. Figure 5.3
shows that the whole robot goes straight without deviation, but the motion of
each leg blurs the center of the body. This blurring was about 1 cm to the right
and left for a movement of about 30 cm by actual survey 16]17].
Figure 5.4 shows the position transitions of each leg, the center of gravity of
the body, and positions of the head and tail, as the amphibious robot turns left.
 143

0.6
-- 1- Left fiont leg
0.5
-Rght fiont leg
z 0.4
Y
-Left rear leg
.s 0.3 -Right rear leg
P
* 0.2 -Head
0.1
-The center of gra*
-Tail
0
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8
Xaxis[m]
Fig. 5.4. Experimental analysis of turning to left
Figure 5.4 shows three inputs of an order to turn left. One turn command
provides a turn of about 30 degrees; therefore, the robot turns 90 degrees in the
figure. A moving leg draws a circle around the center of gravity and moves,
while the center of gravity does not change. Thus, the robot can change
direction while remaining in place. This allows it to work effectively within a
small domain, where other robots do not have the space necessary to change
direction.

References
1. Taro NAKAMURA, Takahiro OMATA, Keisuke IINUMA, Soh FUJII,
“Development of a water strider robot which moves surface on the water,”
JSME Conference on Robotics and Mechatronics, 2PI-A18,2006
2. Keisuke IINUMA, Soh FUJII, Taro NAKAMURA, “Development of an
Amphibious Robot based on a Water Strider,” JSME Conference on Robotics
and Mechatronics, lPl-C05,2007
3. Shigeo HIROSE, Hiroyulu IWASAKI, Yuji UMETANI, “The Basic Study of
the Intelligent Gait Control for Quadrupedal Walking Vehicle,” The Society
of Instrument and Control Engineers (SICE), Japan, 1980.
4. Hideki HAYASHI, Toshiyuki KONDO, Koji ITO, “Biomimetic Hexapod
Robot Using Mechanical Compliance,” The Society of Instrument and
Control Engineers (SICE), Japan, 2002.
5. Noriho KOYACHI, Hironori ADACHI, Tatsuo ARAI, Makoto IZUMI,
Takeshi HIROSE, Naofumi SENJO, Ryoji MURATA, “Walk and
Manipulation by a Hexapod with Integrated Limb Mechanism of Leg & Arm
-Configuration Change of Body and Limbs for Dual Arm Manipulation,” The
Robotics Society of Japan, V01.22, No.3, pp.411-421, 2004.
6. Shigeo HIROSE, Yuji UMETANI, “The Synthesis of Basic Motion
Regulator System for Quadrupedal Walking Vehicle and Its Experiments,”
The Society of Instrument and Control Engineers (SICE), Japan, 1979.
7. Seiji MASAKADO, Takayuki ISHII, Kazuo ISHII, “A Gait-Transition with
Stability Secured for a Quadruped Walking Robot,” The Robotics Society
of Japan, Japan, 2004.
 DEVELOPMENT OF AN OMNI-DIRECTIONAL MOBILE
ROBOT BASED OM SNAIL LOCOMOTION

KUNIAKI SATOH
Major of Precision Engineering, Graduate School of Science and Engineering,
Chuo University,
1-13-27 Kasuga, Bunkyo-ku, Tokyo 112-8551, Japan

TARO NAKAMURA
Major ofPrecision Engineering, Graduate School of Science and Engineering,
Chuo University,
1-13-27Kasuga, Bunkyo-ku, Tokyo 112-8551,Japan

In recent years in the field of bio-mechatronics, a robot that imitates a creature’s motion
has been developed. In this study, we present a locomotion robot based on locomotion of
snail. The snail moves by propagating traveling waves from tail to head. If it is possible
to propagate a traveling wave in many directions, an omni-directional mobile robot could
be realized. Since the locomotion mechanism of the snail involves moving a larger area
than in the case with other creatures, it is able to move not only on irregular ground such
as swamps, but also on walls and ceilings. We have developed an omni-directional
mobile robot that makes use of a traveling wave and have confirmed that the robot can
move in many directions using a traveling wave.

1. Introduction
At present, the wheel mechanism is the main mechanism for moving over
ground. However, this has a disadvantage that it is difficult to move freely over
bad ground such as swamps and wasteland. In addition, it is difficult for the
wheel to move in a diagonal direction. Therefore, to realize omni-directional
locomotion, a special tire is necessary such as omni-wheel. However, this has
the disadvantage that the wheel slides axially on a steep gradient.
On the other hand, in recent years in the field of bio-mechatronics, a robot that
imitates a creature’s motion has been developed[’][*].Here, we report a study on
the locomotion mechanism of snail. The snail has an elastic motion in the
direction of movement, and this is called a pedal locomotion mechanism. In this
mechanism, a traveling wave propagates from tail to head. Since the locomotion
mechanism of the snail involves moving a larger area than in the case with other
creatures, it is able to move not only on irregular surfaces such as swamps, but
also on walls and ceilings. Also, because of the mechanism to propagate the
wave in the direction of movement, omni-directional motion may be expected.

144
 145

In this research, we developed an omni-directional mobile robot using the

locomotion mechanism of the snail and studied the basic properties of the robot.

2. The locomotion mechanism of the snail

The common belief is that the snail moves by a traveling wave locomotion
mechanism. This produces a traveling wave in the gastropod’s foot that
propagates in the direction of movement[31.The state of the traveling wave seen
from a section of the gastropod’s foot is shown in Fig.2.1, and the state of the
traveling wave when moving forward is shown in Fig.2.2. The snail has two
kinds of muscles: the anterior oblique and the posterior oblique muscles
(Fig.2.1). These produce a traveling wave that propagates from tail to head
(Fig.2.2). When the traveling wave occurs, the part that is not forming the
traveling wave in the gastropod’s foot remains in contact with the ground. The
ground generates a frictional force with the grounded foot of the gastropod.

Posterior oblique Anterior oblique

Surface of sole Wave Substrate

Movement of wave and foot

Fig.2.1 Diagrammatic section of snail’s footL4].

Traveling wave direction

Propagation direction

Fig.2.2 Propagation of traveling wave of snail15]

146

3. The omni-directional mobile robot

3.1. Overview of the robot

We have developed an omni-directional mobile robot that uses the locomotion
ism of the snail (Fig.3.1). The specifications of the robot are shown in
1. The omni-directional mobile robot is composed of eight units
arranged in a circular shape and each abutting unit is connected by a spring.

Fig.3.1 O m ~ - ~ r e c t i o nmobile
al robot'61

Tab.3.1 Specificationsof the robot

 147

3.2. Unit
A diagram of the unit used in the omni-directional mobile robot is shown in
Fig.3.2. It is composed of two substrate parts (side: 20 mm, depth: 105 mm): the
friction part (side: 25 mm, depth: 105 mm) and a servomotor.
A friction part

Substrate parts

During expansion During contraction

Fig.3.2 Mechanism of the unit"].

Tab.3.2 Specification of the unit

Width (mm)

Weight (g)

The substrate part is always in contact with the ground. In the condition
where the substrate part is grounded, the unit has elasticity in the direction of
movement.
The frictional part is the part that holds the unit stationary when the unit is
elongating. A frictional sheet is attached to the contact area of the frictional part.
Then, this frictional part is connected to the substrate part using parallel linkages.
148

By using a parallel linkage, during elongation, the frictional part connects to

the ground horizontally, and during contraction, the frictional part leaves the
ground horizontally.
We fixed a servomotor to the substrate part, which is connected to the other
substrate part using a crank mechanism.

3.3. ~ e t h o od ~ l o c o m o t i o ~
When the unit is stationery, it adopts the elongation condition from the
servomotor, the hctional part is in contact with the ground and the unit is fixed
on the ground.
On the other hand, when the unit is moving, it adopts the contraction
condition from the servomotor, and the frictional part does not come in contact
with the ground. As a result, because the friction of the unit is small, it is able to
move freely.
As mentioned above, the robot generates a traveling wave by elongation and
contraction of the units and realizes an omni-directional locomotion.
Figure 3.3 shows the locomotion of the robot. The default position is the
condition where all the units are elongated. First, the robot constricts the pair of
units shown as unitl at the same time. Then, the robot constricts the unit2 pair at
the same time of elongating the unitl pair. Later, the robot constricts the unit3
pair at the same time of elongating the unit2 pair and then constricts the unit4
pair at the same time of elongating the unit3 pair. Finally, the robot elongates
the unit4 pair and we call this complete cycle 1 period. Because of this
movement, the robot generates the traveling waves shown in Fig.3.3 and moves
forward in the advance direction. To change the advance direction, the robot
changes to a different unitl pair appropriate for the new direction of travel.
These units constrict first and this changes the direction of propagation of the
traveling wave.
Advance direction

direction

Fig.3.3 Movement of the robot in the advance direction

 149

4. The locomotion experiment

4.1. ~ o n ~ r m a t i experiment
on for traveling wave lo~omotion
First, we experimented to confirm whether the robot moved using the traveling
wave that is the locomotion mechanism of the snail. The gage points in the
center of the units are shown in Fig.4.1. Then, we made the robot move in the y
axial direction and confirmed the traveling wave locomotion. The experimental
result is shown in Fig.4.2. This graph shows the relationship between the time
and the axial displacement of y. The most backward units fiom the direction of
movement constrict first and we found that the constriction of the unit
propagates in the direction of movement. From this result, we confirmed that the
robot moved by the traveling wave.

Gage point 5 Gage point 4

Gagepoint6 \ / Gage point 3

ge point 2
Gage p0"mt 7
\
Gage point 8 Gage point 1

Fig.4.1 Points attached at each unit of the robot

150

d""

250
200
150
100
50
0
-50
-100
-150
-200 b 5 10
I
15
Time (s)
Fig.4.2 Confirmation of the traveling wave

4.2. ~irectioncontrol ~ e ~ i m e n t
We experimented to discuss whether the robot moved in the expected
direction. In experiment, as shown in Fig.4.3, the gage point is at the center of
the robot. Next, we made the y axial direction the 0" direction and made the
robot move clockwise in the 45", 90", 135" and 180" directions, respectively.
The experimental results are shown in Fig.4.4. This shows the track that the
robot moved to and the ideal track. The results suggest that the robot moved in
the direction of travel with a reasonable accuracy. However, in the 45" and 135"
directions, the movement track of the robot is slightly off the ideal track. It is
thought that the robot has come off the movement track because the fictional
forces are different in the elastic units on either side of the robot.

Fig.4.3 The mnning experinlent

 151

- Movement track

Displacement to the x axial direction (mm)

Fig.4.4 Track of the robot

5. Conclusions
In this research, we have developed an omni-directional mobile robot using the
locomotion mechanism of the snail and have reached the following conclusions.
* We constructed an omni-directional mobile robot using the locomotion
mechanism of the snail and moved the robot on a flat surface.
We moved the robot and confirmed that it moves using the traveling wave
locomotion.
* With the locomotion experiment, we confirmed that the robot moved in the
expected direction with reasonable accuracy.
152

References

1. Brian Chan, N.J.Balmforth, A.E.Hosoi, Building a better snail: Lubrication

and adhesive locomotion, PHYSICS OF FLUIDS, 2005
2. Bai Chen, Yinsheng Zhou, Daqiang Gu, Jie Zhong, A Novel Snail-Like
Micro Robot, ICIMA, China August 2004
3. Ryousen FUJIHARA, Hirohisa MORIKAWA, Shun-ichi KOBAYASHI, The
Mechanism of Pedal Locomotion of Gastropod, The Japan Society of
Mechanical Engineers, JAPAN, vo167 No.658 (2001-6)
4. H.D.Jones, Circulatory pressures in Helix pomatia, L.Comp,Biochem.Physiol,
39A, 2894 1971)
5. Ryousen FUJIHARA, Hirohisa MORIKAWA, Yuya HUKUYA, Hiroshi
SAKAI, Shunichi KOBAYASHI, Pedal-Like Locomotion Mechanism Modeled
on Pedal Crawling of Terrestrial Gastropod, The Japan Society of Mechanical
Engineers, JAPAN, V0170 No.695 (2004-7)
6. Shohei IMAI, Kuniaki SATOU, Taro NAKAMURA, Development of an
Omni-Directional Mobile Robot Based on Snail Locomotion, JAPAN, JSME
Conference on Robotics and Mechatronics Conference, 1P1-C06,2007
7. Yuki MITAMURA, Syohei IMAI, Taro NAKAMURA, Development of a
traveling wave robot based on snail locomotion, JAPAN, JSME Conference on
Robotics and Mechatronics Conference, 2P 1-A19,2006
GAIT PARAMETER ADAPTATION TO ENVIRONMENTAL
PERTURBATIONS IN QUADRUPEDAL ROBOTS

E. GARCIA*, J. ESTREMERA, P. GONZALEZ DE SANTOS and M. ARMADA

Industrial Automation Institute - CSIC,
28500 La Poveda, Madrid, Spain
*E-mail: egarcia9iai.csic.es
w w . iai. csic. es/users/egarcia

Quadrupedal robots working outdoors are very slow robots prone to tum-
ble down in the presence of perturbations. This paper presents a novel gait-
adaptation method that enables walking-machine gaits to autonomously adapt
to environmental perturbations, including the slope of the terrain, by finding
the gait parameters that maximize robot’s dynamic stability. Experiments with
the SILO4 quadruped robot are presented and show how robot stability is more
robust when the proposed approach is used for different external forces and
sloping terrains.

Keywords: Quadruped robot; Gait adaptation; Dynamic stability margin.

1. Introduction
Walking robots designed for field and service applications usually perform
statically stable gaits, which have been designed to optimize a static sta-
bility margin. Even when a dynamic stability margin is used to control the
robot’s motion, the gait pattern is a statically-stable one. As a result of
this contradiction, the robot will still tumble down when confronted with
any perturbing effect. Although an accurate dynamic stability margin mea-
sures robot stability, unstable situations are merely observed, not avoided,
unless the gait pattern is modified. Existing dynamic gaits like the trot
and the gallop require light, simplified robot designs with elastic actuators,
which differ greatly from the heavy-limbed machines1’2used for field and
service applications. Therefore, statically stable robot gaits should be mod-
ified based on a more suitable dynamic stability margin to improve gait
control.
Only a few researchers have attended to partially solve this problem,
achieving adaptation to uneven terrain, yet they did not consider robot

153
154

Leg workspace

Yc,,mah Yc+,min
!............................... !
1
x,
yc2,max Yc2,min
!...............................i

j (XcZ.Yc2) i

Yc3,max Ya,min Yotrnax Yd,min

Fig. 1. Outlining of gait-parameter adaptation for a quadruped in the body-motion

phase of a two-phase discontinuous gait

dynamics nor environmental perturbation^.^^^ This paper presents a new

method for gait adaptation to the environment based on the Normalized
Dynamic Energy Stability Margin, S N L I E S M . ~ method maximizes the
The
dynamic stability margin during the transfer and support phases and thus
we get a more robust gait against external perturbations. Section 2 de-
scribes the gait-parameter adaptation approach. Experiments using the
SILO4 walking robot (see Figure 5) are described in Section 3 to show
the improvement of the gait-parameter adaptation method when external
forces are applied.

2. Gait Adaptation
Let us assume a quadruped robot walking along the direction of its longi-
tudinal body axis. Let us define an external, fixed reference frame {x,y, z }
and a body-fixed reference frame, {x,,yc, z,} centered at the robot’s center
of gravity (CG) so that robot motion is along the x, axis and the z, axis
is orthogonal to the body plane. The terrain can be inclined, forming an
angle 8 with the x-axis (pitch angle) and an angle B with the y-axis (roll
angle). The gait optimization procedure is divided into the following two
phases:
 155

2.1. First phase: Leg-stroke and foothold calculation

In this phase, foot and body CG positions in the body plane and the leg
stroke that optimize the gait are determined. Figure 1 outlines a top view of
leg workspaces, footholds and leg stroke for a quadruped during the body-
motion phase in a two-phase discontinuos gait.5 The four legs are in support,
thus propelling the body forward. The initial position of the robot has been
plotted in thick, solid line, while the final position of the robot has been
plotted in thick, dashed line. Let us name xci and yci the 2,-coordinate
and yc-coordinate respectively of foothold i referring to the body frame
{xc,yc, 2 , ) just before the body-motion phase starts. The CG trajectory
has been also plotted ( x ~ Gy, c ~ )finding
, the maximum-stability position at
the middle of the trajectory, as explained before. The workspace of leg i
is delimited by xcz,,, , xCimzn , ycimaz, yczmzn.Notice that the length of the
body motion is half the leg stroke, that is, s = 2d. Therefore, maximizing s
is achieved by maximizing d, which can be expressed as a function of front
legs' footholds:
d(XC1,xc2) = xc2 - Z C 1 . (1)
The two rear legs' footholds can be expressed in terms of the two front
legs' footholds:

xc3 = xc3,zn + xc2 xc1

- (2)
5 c 4 = XC3,,, + 2xc2 - 2%. (3)
The stroke pitch in the yc-axis direction, Py,provides the following
relation to foothold yc components:
Ycz = Ycl - P y (4)
Yc4 = Yc3 - py. (5)
To obtain the optimized gait, xcl, ycl, xc2,'yc3, X,G and Y,-G have to be
obtained so that S N D E S M is maximum at the middle of the body-motion
phase. This condition will equal the probabilities of losing robot stability
when a rear leg is lifted or when a front leg is lifted, thus decreasing the
overall probability of tumbling. Therefore, S N D E S Mshould be expressed as
a function of footholds and CG position in the body reference frame, that
is (see Ref. 6 for a more detailed explanation):

S N D E S M = ~ ( ~ c l , Y c l Yc31XcGr
~ ~ c Z ~YcG, Z c G ) . (6)
Taking into consideration that in this first phase of the optimization
process the CG height is considered constant, Z,G = ho. S N D E S Mhas
156

........................ 0.4 ....................................... .................... SN&M

0.2

-- 0

-0.2

-0.4

-0.6

-0.5l ' I
-20 -10 0 10 20 -0.
40 -10 0 10 20
(dw) (dw)
(a) (b)
Fig. 2. Results of first-phase gait optimization for sloping terrain (a) Effect of pitch
angle, (b) Effect of roll angle.

to be maximum when the body CG is located at the middle of the CG

trajectory; therefore the objective function is:

However, another objective function exists, because the leg stroke must
be maximum too:

J2 = d(XC1, x c 2 ) (8)
with the following constraints:

XClmin I xc1 I Xclmas (9)

Xc2& I xc2 I 5c2mar (10)
Ycl,,, I Ycl 5 Ycl,,, (11)
YC3,in I Yc3 I YC3,,, (12)
4Xc1 - 3Xc2 2 <I - ~clmas (13)
Constraints (9) to (12) are given by the leg-workspace limits, while con-
straint (13) avoids leg-workspace overlapping, where & 0 is a constant. >
To solve the multi-objective problem, the &-ConstraintMethod is used.7
The problem has been solved numerically for the SILO4 robot in an iterative
manner starting from an initial estimate of foot positions 21 and x 2 . Figures
2 and 3 show results of the first phase optimization for the medium-sized
 0.6

0.4

0.2

0
--0.2
E

-0.4

-0.6

-200 -100 0 100 200 -0.

- ~ 0 0 -100 0 100 i
Fx (N) F, (N)

Fig. 3. Results of first-phase gait optimization for external forces (a) x component of
force, (b) y component of force.

quadruped robot SIL04. Figure 2 shows footholds, leg stroke and maximum
S N D E ~obtained
M for a robot walking on sloping terrain. As the terrain
angle increases the robot has to modify foothold x, and yc components
linearly to decrease the leg stroke (see d) and increase S E g E S M . Similar
results have been found at the first phase optimization when external forces
are applied (see Figures 3(a) and (b)). Note that the direct consequence of
the linear reduction of leg stroke for increasing perturbations is a restriction
in leg workspace. Thus, the bigger the disturbance, the bigger the reduction
of leg workspace.

2 . 2 . Second phase: CG-height calculation

When, as a result of the first phase of the optimization approach the leg
stroke is reduced to adapt the walk to a stepped ground or an external force
and enhance stability during the body-motion phase, stability is reduced
during the transfer phase. This problem can be solved by reducing the CG
height. Thus, the following condition is stated:

3 ( h )- S g D E S M = 0 (14)
where SkDESMstands for a constant value of the gait stability margin in
nominal conditions, that is, over flat terrain when there are no external
disturbances. The function F ( h ) is the result from the first phase of the
158

(4 (b)
Fig. 4. Results of second-phase gait optimization for (a) sloping terrain, (b) external
force.

Fig. 5. Quadruped robot SILO4 pulling a 50-N load

optimization procedure, where every gait parameter has been obtained.

Therefore, the only variable at this phase is CG height (h).This problem
is a nonlinear equation of a single variable that has been solved by least-
squares optimization.
Figure 4 shows results of CG height for the SILO4 walking (a) in sloping
terrain and (b) when external forces are applied.

3. Experiments with SILO4

The improvement of gait stability obtained by the use of the gait-
optimization method herein proposed has been proved experimentally using
the SILO4 robot5is shown in Figure 5. A real service application has been
simulated. The SILO4 robot walks while it pulls a load, thus being affected
by an external force'?l = -50N. As a consequence of the perturbation
the robot finally tumbles down. Then, the experiment is repeated using the
gait-optimization approach, which modifies the gait pattern for the -50-N
 159

0 5 10 15 20 25 30 35 40 45
Time (s)

(4

0 5 10 15 20 25 30 35 40 45
Time ( 5 )

(b)
Fig. 6. Time evolution of S N D E S Mfor the SILO4 robot walking against a 50-N dis-
turbance. (a) Standard gait, (b) Optimized gait using the proposed approach.

external force. In this experiment the robot pulls successfully the same load
that made the robot tumble down in the previous experiment.
Figure 6(a) shows the evolution of the stability margin during the first
experiment, when the SILO4 robot walks against the disturbance using a
standard two-phase discontinuous gait. The stability margin drops during
the transfer phase of the rear legs (because the destabilizing effects of leg
transfer and force disturbance combine). Let us compare Figure 6(a) with
Figure 6(b) which shows the evolution of the stability margin during the
second experiment, when the SILO4 robot walks using the results of the
gait optimization. As a result of adapting the gait to a -50-N force, the
stability margin is symmetrical for the transfer phases of both the rear and
the front legs. As a result of the adaptation to a -50-N force, the gait is
more stable than in Figure 6(a), increasing the gait stability margin by
70% in this experiment. As shown in the gait optimization results, the use
of the proposed gait-parameter adaptation approach increases the robot’s
robustness to external disturbances from a 50-N load up to a 200-N load,
that is, four times the perturbation in this particular robot.

4. Conclusions
A new gait-adaptation method has been presented in this paper to cope
with external perturbations. The method enables quadrupedal gaits to
160

adapt to the slope of the terrain and external forces by finding the gait pa-
rameters that maximize robot stability. The resulting footholds of the gait
optimization approach are linear functions of external forces and terrain-
inclination angles. This reduces the computation time and enables the ap-
plication of the gait-optimization method in real time.
Experimental results with the SILO4 robot have shown how robot stai
bility is enhanced when the proposed approach is used t o counteract ex-
ternal forces. The improvement in gait stability margin when applying the
proposed gait-adaptation method t o the SILO4 robot has been proved t o
be 70 percent for a -50-N external force. Also, the resulting gait increases
the SILO4 robot’s robustness to external disturbances from a 50-N load
up to a 200-N load, that is, four times the perturbation in this particular
robot. The proposed method has been shown t o be of major relevance for
the use of walking robots in field and service applications, where robots and
environment interact.

Acknowledgements
This work has been funded by the Spanish Ministry of Education and Sci-
ence through Grant DPI2004-05824.

References
1. P. Gonzalez de Santos, E. Garcia, J. Estremera and M. Armada, International
Journal of Systems Science 36, 545 (2005).
2. T. Doi, R. Hodoshima,Y. Fukuda, S. Hirose, T. Okamoto and J. Mori, Journal
of Robotics and Mechatronics 18,318 (2006).
3. D. Wettergreen and C. Thorpe, Developing planning and reactive control for a
hexapod robot, in Proc. I E E E Int. Conf.Robotics and Automation, (Atlanta,
Georgia, 1996).
4. H. Tsukagoshi and S. Hirose, Intermittent crawl gait for quadruped walking
vehicles on rough terrain, in Int. Conf. Climbing and Walking Robots, (Brus-
sels, Belgium, 1998).
5. P. Gonzalez de Santos, E. Garcia and J. Estremera, Quadrupedal Locomo-
tion: A n Introduction to the Control of Four-Legged Robots (Springer-Verlag,
London, 2006).
6. E. Garcia and P. Gonzalez de Santos, I E E E Transactions on Robotics 22, 1240
(2006).
7. V. Chankong and Y. V. Haimes, Multiobjective Decision MakingTheory and
Methodology (Elsevier, New York, 1983).
8. Industrial Automation Institute, C.S.I.C., Madrid, Spain, The SILO4 Walking
Robot, (2002). Available: h t t p : / / w .iai.csic.es/users/silo4/.
 INTELLIGENT SPIDER* WALKING ROBOT FOR ROUGH
TERRAIN

MICHAEL MCCREADY, LIQIONG TANG AND GURVINDER VIRK

Institute of Technologyand Engineering

Massey University
New Zealand

This paper presents the design of a legged walking robot capable of traversing rough
terrain that would be difficult to negotiate using wheeled vehicles. The body of the robot
looks like a spider but there are only six legs in order that various insect gaits can be
studied. Each leg is driven by a worm gear which allows each leg to make a vertical
movement individually thereby realizing easy control of the leg height. The design
incorporates two cam systems to synchronize the horizontal movement of the legs on
both sides of the robot’s body. Each cam is driven by a stepper motor and only controls
the three legs on one side. Such a design allows steering of the robot body to be
controlled by the speed difference of the two motors. The mechanical system of the robot
and the motor driving systems have been built. An intelligent control system is still under
development. On completion, it is expected that the robot will be smart enough to walk
around and collect useful information for its applications. The paper presents the details
of the mechanical system, the walk patterns, and the driving system. The intelligent
features to be added to the robot are also discussed.
* SPIDER - Special Purpose Insectoid Drive Experimental Rover

1. Introduction

The aim of this project is to design and build a legged walking robot, much like
a spider, that would be capable of traversing terrain that would otherwise be
difficult or unstable for a wheeled or tracked vehicle. Unlike other types of
systems, walking robots shown in Figure 1 are more complicated to design due
to their relative complexity, as they normally require complex mechanical
systems to support the walking algorithm [1],[2]. Most existing walking robot
requires many motors to realize the movements and are often suited for a sole
task. Without major modification to the design, it is difficult to use for another
job. Therefore a general purpose walking robot, which could be easily adapted
depending on the job required, is of interest. This project emphasises simplicity
in the design with mechanical components that are easy to manufacture. It
incorporates modularity in the design concept so that the later development of
add-on units such as vision, sound, and other mechanical and electronic control
units which would use their own processor can be added to enhance the control
and movement commands to the controller for the legs.

161
162

NASA Spider-bot- Space Exploration Plustech Walkincr Forest Machine - Forestrv

Figure 1. Walking robots

The robot presented is a six-leg walking robot as shown in Figure 2. Such a

design gives better proportions o f width to length and also reduces the number
o f components and weight. The body of the robot are built by layers, Each layer
has a base plate and all the base plates have the same profile and dimensions.
Such a body structure provides an easy add-on platform for future developments
to be added. A cam system is employed to control the robot’s forward and
backward leg movements. Leg lift is individually controlled through six small
motors, which makes the robot able to walk on uneven and sloped surfaces. The
control system is under development. The system will provide an easy platform
for future development to integrate with new software and hardware and add-on
new ~ c t i o n a l i ~toe sthe robot.

Figure 2. Massey’s SPIDER Robot

 163

2. Mechanical Driving System

The SPIDER robot has six legs. It is possible to let three legs on the ground to
support the body while moving the other legs forward or backward. This project
prefers using four legs to support the robot. There are different methods and
algorithms to control the robot legs to achieve different walking gaits [3], [4],
[ 5 ] . After a good study and research on leg movement, the decision was made
to incorporate a cam system to realise the horizontal leg movement. The cam
system is required to make four of the six legs, two on each side, stay on the
ground at a time giving maximum support to the robot body and also obtain a
good stability while two of the legs move forward or backward. A leg lifting
system is designed to work closely with the cam system. Only until the two legs
are lifted, can the two legs move in the desired direction. The cam system
developed consists of six cams, three on each side of the robot. All six cams
have identical profile that has been specifically designed, which moves the leg
forward within the first third of a revolution. During the other two third
revolution, the leg stays on the ground as shown in Figure 3. With a 120" offset
between the three cams on one side of the robot, it is possible to have two legs
on the ground at a time. Combining the legs on both sides, the cam system is
able to move two legs forward while the other four legs stay on the ground to
support the robot body. With different offset between the two sides, it is
possible to make any two legs as a pair. This provides more choices for
studying the moving algorithm.

70.00

60.00
h

50.00
v
+
5 40.00
E
30.00
-c1Q 20.00
.-
P
10.00

0.00 i
0 60 120 180 240 3M) 360
Rotation (Degrees)
Takes K rotation to lift leg fonvard
Then K rotation to pull robot forward
Cams 120" out of phase so four legs always on ground
Figure 3. Cam profile and offset
164

The cam driving and linkage system employs two stepper motors where each
motor controls the three cams on one side of the robot. The mechanical
arrangement is presented in Figure 4. Through the bevel gears, each stepper
motors is able to drive the three cams underneath the second layer board which
linked to the three robot legs. The cam profile is designed based-on the cam
driving system and the dimension of the linkage. The distance required to move
the leg forward and backward is also one of the main inputs for the cam profile
design.

Figure 4. Cam driving and linkage system

3. Leg Structure and Control

The robot leg structure, apart from realising forward and backward
movement, is also required to be able to make the leg move up and
down. The leg design is made on the basis that the leg components must
be easily manufactured with the available tools in a typical engineering
workshop. All the leg components have simple shape and are easy to be
made and assembled. Since this is the first prototype it was preferred
not to use bearings in the joints in order to reduce the cost, but add
washers to separate the surface areas during movement, and use nylon
at the waist joint to prevent grinding.
 165

Figure 5. Robot leg assembly

In the leg assembly shown in Figure 5 the pivot point from the waist is
designed to have an angle of 10" from the vertical. This makes the leg
always be splayed at an angle of 10" to make it more stable. To
complement the lifting movement, the motor is placed above the waist
and the nut mounted in the lower section of the thigh, close to the knee.
This means the motor's output torque would be utilised to a much
greater potential.

A test model was made and produced to check the design. The model
was capable of making the required movements. The prototype also
revealed there was too great a stress on the output shaft of the motor.
The modified version incorporates a bearing to the mount to take the
load off the motor. Such a modification worked well and a further five
legs were built to complete the full prototype model. Figure 5 shows the
complete leg design with the key parts highlighted.

The robot leg can move up and down and forward and backward. To
synchronise these movements such as the leg lift must be activated
before any forward movement, a sensing system must be implemented
to give the appropriate feedback to the robot controller. Such a sensing
166

system can also prevent parts from moving outside the limit positions.
Two limit sensors are installed in the leg assembly that would ’catch‘ the
leg if it moved to a particular limit and send a signal to the controller to
stop driving. This also prevents the worm gear winding out of the nut if
the leg was lowering, and the motor mount being crushed if the leg was
lifting up. The mechanical structure of the robot leg provides an ideal
location for the sensors, which makes it possible to use only one sensor
per leg to detect both upper and lower limits, and to keep tracking the
leg position using software. The sensor is mounted in between the two
rails of the upper thigh as shown in Figure 5. Floor detection sensors are
planned to be installed in the robot’s feet to get the feedback when the
leg touches the ground, allowing the robot to position its legs at different
levels on uneven terrain. The robot feet design is another design task.
The floor detection sensors will be installed at the bottom of the robot
feet.

4. Robot Walking Patterns

4.1 Forward and Backward Movement

The cam drive systems synchronise the robot to move forward and backward.
Table 1 is an example to move the robot to make forward movement. As stated
above, the cam system can only make two legs move forward at a time and the
other four will stay on the ground to pull the robot body. Through offsetting the
cams, it is possible to make any two legs as a pair to move forward. Table 1
illustrates the movement of the legs when the cams rotate one revolution. The
symbol “A” represents the legs moving forward and ‘‘d’ means that the legs
stay on the ground to pull the body forward.

Table 1. Robot Forward and Backward Movement

 167

4.2 Lefl and Right Movement

The left and right movement is controlled by DC motors. The six DC motors
mounted on each of the robot legs can move the robot leg individually to
different height. When the robot is under the left and right movement mode, the
cam system is locked. Each move on the left or right will only lift three legs.
The other three are kept on the ground to support the robot’s body. Table 2
illustrates one of the possible ways to make the robot walk to the right.

5. Robot Control System

Two stepper motors are employed to drive the robot’s legs move forward and
backward and six DC motors for individually lifting the legs up and down. The
motor driver circuits for the two stepper motors are built using two ICs, L297
and L298. Additionally there is an on-board chopper circuit, designed to take
168

current sensing signals from the driver, and a reference voltage input, and
generate a PWM output to the driver. This allows the use of a higher voltage
supply than the motor is rated at, and keeps a higher current through its
windings to maximise the output power. The six DC motors are controlled
through a board made using L293 driver ICs. The prototype control system uses
a microcontroller from Silicon Laboratories. The 805 1 microprocessor has the
features the project required. After integrating the motor drive circuit boards
and the microcontroller, the robot is able to move each of the legs up and down
and drive forwards and backwards by the cams. Further developments are still
under way. Once the project is completed, the SPIDER should be able to move
around with a wireless user interface.

6. Conclusion
The spider robot presented integrates mechanical design, electronic
circuit board design, control algorithms development, and programming.
It produces a good platform for study and testing different walking
patterns and control algorithms. The layered mechanical structure also
makes it easy for future add-ons and to integrate new electronic,
mechanical and software modules.

References
1. P. G. de Santos, E Garcia and J Estremera, “Quadrupedal locomotion : an
introduction to the control of four-legged robots”, Springer (2006).
2. S. Shoval, E. Rimon and A. Shapiro, “Design of a Spider-Like Robot for
Motion with Quasistatic Force Constraints”, Proceedings of the 1999 JEEE
International Conference on Robotics & Automation, Michigan (1999).
3. John J. Craig, “Introduction to robotics: mechanics & control”, Addison-
Wesley Pub. Co. (1986).
4. R. Quinn, G. Nelson, R. Bachmann, D. Kingsley, J. and R. Ritzmann,
“Insect Designs for Improved Robot Mobility”, Proc. 4 t h Int. Con$ On
Climbing and Walking Robots, Berm and Dillmann eds., Pro$ Eng. Pub.,
69-76 (2001).
5 . R. Bachmann, G. Nelson, R. Quinn, etc., “Construction of a Cockroach-like
Hexapod Robot”, In Proc. Sixth IASTED Int. Con$ on Robotics and
Automation, Banff, Canada, pp. 22-27 (1998).
 ~ ~ ~ A T I SENSORS
C S , AND CONTROL OF THE FULLY
~ U T O ~ A T EFACADE
D CLEANING ROBOT SIRIUSC FOR
THE F ~ U ~ O F E H ERA D Q U ~ T E R SBUILDUNG, ~ ~ I C H

NORBERT ELKMANN, MARIO LUCKE, TINO KRUGER, DIETMAR KUNST,

THOMAS STURZE
Robotic Systems, Fraunhofer Institutefor Factory Operation and Automat~on,
Sandtorstrasse 22, 391 06Magdeburg, Germany

The Fraunhofer Institute for Factory Operation and Automation IFF has developed the
automatic facade cleaning robot SIRIUSc for use on the Fraunhofer-Gesellschafts
headquarters, a high-rise building in Munich, Germany. The building has a height of
80m, its facade an area of 4000 mz. Apart from the robot that moves along and cleans the
facade, the complete, fully automated system consists of a fully automated gantry that
secures, supplies energy to and above all positions the robot. Part of the project involved
completely automating a standard gantry, which is an integral part of the complete facade
cleaning robot system. This article presents an overview of the significantbasic functions
of the robot and the gantry, emphasizing the kinematics, the control and sensor systems
for navigation and the cleaning sequence that employs the extensive fully automatic
functions of the robot and gantry.

1. Introduction
The complete faGade cleaning robot system consists of more than the robot
alone (Figure 1). The rooftop gantry positions the robot at the top of the facade
on every pane path to be cleaned.

Carries cable w i n ~ l i

Kinemdtics for surface motion and

ncgotiation of obbtdLIes
Control $ystern,wisorb, navigation,
operator interfilcc
Cledillng tool
1 igure 1 Roofiop gdniry and fxdde cleaning robot dlop the building

169
170

The robot then descends the facade vertically and cleans when it ascends.
Four cables connected to the gantry atop the building safeguard the robot
against falling. Since the cables must be taut to ensure the robot is secure, the
cables also used to position the robot as the winding and unwinding of the
winch on the rooftop gantry moves it vertically along the building as well as to
bear the load of the robot. Cables transmit data and supply power too. The robot
is one of the first fully automatic systems of this kind used in a public setting.
The robot weighs 450 kg, the gantry 5000 kg.

2. Rooftop Gantry Crane

The gantry (Figure 2) possesses three degrees of freedom relevant for
positioning the robot on the facade: Movement along the rails and two rotary
cantilever arms.
Measuring systems help position the gantry on the roof for its start on the
pane path to be cleaned. The gantry has two asynchronous drives to move along
the rails. Two encoders determine the gantry’s position. The gantry is positioned
by a continuous controller that reads the pane path positions taught out of a data
module of the control system. Another redundant measuring system has been
installed because of the slip.

pane path

Figure 2. Gantrys degrees of freedom to position and deposit the robot on the facade
 171

When the encoder target position is reached, the transponders mounted on

each of the pane path target positions must be matched with the transponder
system. If the transponder assigned to each pane path can be read out
successfully, the target position is reached and the positioning of the cantilever
arm can begin. Otherwise an error code is generated and the start on the pane
path must be repeated.
A synchronous movement of the two rotary cantilever arm drives sets the
robot on the facade. To acheve a high level of security when positioning the
robot, its orientation parallel to the facade during motion must be guaranteed. A
cascaded controller in the gantry control system ensures this.
A particular challenge is positioning the robot and thus the gantry on the
corners of the facade since the rails executes a curve and do not run parallel to
the facade. Here the gantry’s traveling mechanism (translation) and the
cantilever arms (rotation) must run sequentially synchronized to prevent any
collision of the robot with the facade. In order to reduce the robot’s distance to
the facade and be able to deposit the robot on it, the gantry must proceed along
the rails. This simultaneously influences the robot’s orientation plane-parallel to
the facade negatively. To simplify the positioning process, the traveling
mechanism is guided in a second step to the window pane target position and
then the cantilever arms are moved again so that the robot reaches its ideal
position (distance of approximately l O O m m and parallel to the facade). All
operations are fully automated. The gantry is a standard piece of equipment that
was completely automated in the project.

3. SIRIUSc Kinematics
Above all, the modular kinematics (Figure 4) ensures the robot remains in
constant contact with the facade and is able to navigate a multitude of typical
obstacles and move quickly along a facadL€hekinematics is based on a
structure of two pairs of linear modules, the so-called “advanced sliding module
mechanism” [8,9]. Two linear modules constitute one pair that performs the
same linear movement, thus ensuring secure and stable contact with the facade.
Servo drives move each of the outer or inner sucker units to the next
position on a pane within a frame. The linear modules with their suckers are
positioned in such a way that the suckers are located above or below the
horizontal pane frames when docking onto the facade. Once a sucker unit has
suctioned on, the drive’s brake is deactivated and the other unit’s suckers are
released and retracted. The winding and unwinding of the securing cables on the
gantry lift winch produces the robot’s upward and downward motion on the
172

facade. The alternation of the outer and inner sucker units each with 2x3
suckers produces the walking motion of the robot. While the robot is lifted or
lowered to the desired position, the other activated linear module pair is moved
to the next free position.

Figure 3. Kinematics and navigation sensors of SIRIUSc

Eddy current sensors mounted on the top and bottom of the robot’s body
detect and store the positions of horizontal pane frames.
One of the requirements for outdoor operation in high winds is that the
robot is able to correct its direction of motion, should it drift a little off course.
To this end, the inner pair of linear modules is tilt adjustable to enable small
steering movements to keep the robot on a straight path. Two drift scanners
have been mounted on the robot to control drift compensation. The scanner’s
job is to detect the robot’s position in relation to pane fiames (vertical pane
jambs) in order to systematically control the inner linear modules. A drift
scanner consist of 2x4 eddy current sensors (spaced SOmm), which a controller
assembly switches off and on in succession. This is necessary since the eddy
 173

current sensors would otherwise interfere with one another. Depending on the
robot’s direction of movement and which eddy current sensors detect the
metallic pane frames, the pneumatic drift compensation drive turns the inner
linear guide right, to the middle or left.

drift
variation

*I linear modules
_.-

Figure 4.Schematic diagram of the detection of drift variation

The pane frames are made of various materials and profiles. This represents
a special challenge. In extensive testing with various types of sensors, only
novel eddy current sensors with a detection distance of 0-5Omm were
sufficiently reliable. None of the optical or other sensors were sufficiently
reliable under the given conditions (rain, reflected sunlight on the panes).
174

4. Control System and Navigation

The heart of the system is its control system, which receives and combines
sensor data and operator instructions to generate robot actions. Selected for it is
stability and modularity, a programmable logic controller (PLC) is used for the
robot and the gantry control system. The PLC is on board and controls the entire
system. It synchronizes the walking mechanism with the trolley and cleaning
head. All robot motions and actions are fully automated. The robot was
programmed modularly so that it can be transferred to a large number of
different facades with minimal reprogramming effort. The robot does not start
out with information on all the obstacles it will face in its path. Rather, it keeps
track of the surface and obstacles currently under it. Sensors identify and
measure obstacles and window frames. The PLC then generates the appropriate
step lengths for the robot to successfully walk over obstacles and frames. The
PLC ensures that vacuum suckers directly over an obstacle are not engaged
while simultaneously maximizing the number of vacuum suckers in contact with
the facade at any given moment. Like SIRIUS, the cleaning head does not
require detailed information on the surface of the facade. Rather, the cleaning
system has its own sensors that detect obstacles and end positions. The sensor
signals are also incorporated in the onboard PLC program, which then uses the
information to generate the necessary cleaning head movements.
I

Figure 5. Control system concept

 175

The only manual input information the robot requires before starting off on
a given surface is end point data such as the height of the structure or building,
off limit zones and the maximum length of obstacles. The robot automatically
detects any other necessary surface information during operation. An operator
master display is also located in the building so that an operator can monitor the
robot’s progress. A remote maintenance module allows downloading the robot’s
status and sensor data, uploading new program modules and executing simple
operator commands such as the motion commands for a pair of linear modules,
all over the Internet.
Little knowledge about the general structure of a building’s surface is
needed before robot movement can be generated. The input data includes end
positions, moving distances and path characteristics. This a priori data is
supplemented by online sensors that detect the facade surface and search for
possible obstacles. In addition to identieing obstacles, the external sensor
technology also corrects the direction of motion, Sensors detect where the robot
must deviate from a path, e.g. girders or window and panel seals. The robot
control system communicates with the building control system, making sure all
windows are closed in areas being cleaned.

:h
176

5. Conclusion
The Fraunhofer IFF has developed a fully automated facade cleaning robot for
the headquarters of the Fraunhofer-Gesellschaft in Munich. SIRIUSc has been
developed for vertical facades and consists of the main components robots
mechanics and kinematics, rooftop gantry, sensor systems to detect facade
shape, frames and obstacles, control technology and navigation, power supply
and integrated cleaning unit. SIRIUSc was delivered to the facility management
of the Fraunhofer-Gesellschafts headquarters building in 2006. The facility
management staff is now able to clean the facade without any technical support
from the researchers at the Fraunhofer IFF.

References
1. S. Hirose, K. Kawabe: Ceiling Walk Climbing Robot Ninja-11. Proceedings
of CLAWAR 1998, First International Symposium on Mobile, Climbing
and Walking Robots, Brussels 26.-28. October, 1998, pp. 143-147
2. M. M. Cusack, J. G. Thomas: Robotics for the Inspection of Vertical
Surfaces of Buildings and Structures. In 25th ISIR, pp. 287-295.
3. N. Elkmann, T. Felsch, M. Sack, T. Boehme: Modular Climbing Robot for
Outdoor Operations. Proceedings of CLAWAR 1999, Second International
Conference on Climbing and Walking Robots, Portsmouth 13.-15.
September, 1999, pp. 413-419
4. N. Elkmann, T. Felsch, M. Sack, T. Boehme, J. Saenz: SIRIUS: Modular
Climbing Robot for Facade Cleaning and Other Service Jobs. International
Conference on Field and Service Robotics FSR 200 1, Helsinki
5. N. Elkmann, T. Felsch, M. Sack, J.Saenz, J. Hortig: Innovative Service
Robot Systems for Facade Cleaning of Difficult-to-Access Areas.
International Conference on Intelligent Robots and Systems IROS 2002,
Zurich
6. N. Elkmann, Lucke M., Kriiger T., Kunst D, Stiirze T.: SIRIUSc: Fully
Automatic Facade Cleaning Robot for a High-rise Building in Munich,
Germany. ISFUROBOTIK 2006, Munich
 ON FOUR LEGS TOWARDS FLEXIBLE AND FAST
LOCOMOTION

C. KARA, C. HECKHOFF, T. BRANDT, D. SCHRAMM

Chair of Mechatronics, University of Duisburg-Essen,
47057 Duisburg, Germany
E-mail: { kara, heckhoff, brandt, schmmm} @imech.de
www.imech.de

The breathtaking development of electronic components is one key factor in

the fast evolution of mobile robots. Depending on the mission, robots might
operate autonomously or in cooperation with an human operator. In either
case the locomotion of the robot is a key element. In general, legged systems
are more flexible compared t o wheel based systems and can be deployed in
impractical terrain. Quadruped locomotion has the particular advantage to
feature static as well as dynamic walking. Compared to n > 4 legs quadruped
motion can be very agile, which is impressively demonstrated by cheetahs being
the fastest mammal runners. On the other hand, quadrupeds feature static
walking, which is beneficial for large scale walking machines such as legged
excavators. In this paper a static free-gait pattern for quadrupeds allowing
interaction with an human operator is discussed. Since experiments with large
scale walking machines are very expansive, a scaled prototype is presented.
However, the scaled prototype differs from the original system in many details.
To guarantee the portability of the free-gait algorithm, the scaled workspace
of the feet of the walking machine is proposed as criterion.

Keywords: motion generation, gait pattern, free-gait, quadrupedal walking,

scaled prototypes, walking machines ALDURO and ADONIS

1. Introduction
Mobile robots evolve very fast. Sensors, actuators and electronic control
units are among the key elements promoting this development. However,
mobile robots are mainly characterized by their way of locomotion. The
most general classification are wheel based and leg based systems. Wheel
based systems allow a very energy efficient and fast motion, but are not
applicable in impractical terrain. Here, legged systems are beneficial.
Systems with n > 3 legs are particular flexible. They can walk dynam-
ically, but besides, walk statically and, if indicated, stand still. However,

177
178

focusing on dynamic motion the number of legs should be limited. For n > 4
legs the additional mass of the system generally reduces the dynamic per-
formance. This observation is affirmed by the cheetah being the fastest of
all land animals. It can reach speeds up to 110 km/h in short sprints up to
460m and it is able to accelerate from 0 to 100 km/h in 3.5s.
For these reasons quadruped systems seem to be an interesting ap-
proach to flexible as well as to agile locomotion. As a first step towards
quadruped locomotion including static as well as dynamic walking, this
paper addresses a static free-gait pattern. Here, emphasis is put on flexi-
ble motion. Hence, this pattern does not rely on precomputed motion se-
quences. The gait can easily adapt to environmental changes or commands
of an human operator. Besides that, the control software is developed in
a way, that it is generic for any quadruped walking machine. Due to the
characteristic of the free-gait pattern, the criterion of portability is the
workspace of the feet. This is shown by means of two quadruped walk-
ing robots ALDURO' (~mphiboreptilomorphicallyLegged and Wheeled
-
Duisburg Robot, see Fig. 2) and ADONIS (ALDURO Demonstrator with
Onboard Navigation and Intelligent Step Generation, see Fig. 3).

2. Motion Control Software for Generic Quadrupeds

Thinking of a generic system includes the conception of a flexible software.
This issue was one of the main interests during the development of AL-
DURO. The result is a division of the control software into three main
layers as shown in Fig. 1.2

Fig. 1. Structure of proposed control software

 179

Independent of the particular hardware in use, the MotionGenerator

determines the motion of the central body, decides which foot performs the
next step and derives the velocities for this foot according to velocity and
direction demanded by the operator. This procedure is independent of the
hardware due to the abstraction of sensor data and control inputs. The
latter is done by the Controller module that transforms the given feet and
central body velocities into actuator velocities using the inverse kinematics
of the robot and, by comparing the actual and the desired actuator veloc-
ities, determines the actuator signals. In case of ALDURO the actuator
settings are servo valve voltages. In contrast, the ExtendedSensorics module
transforms the sensor data into dimensionless and robot structure indepen-
dent quantities. These two modules are combined into the second layer, the
HardwareAbstraction. The remaining third layer represents the interface to
the particular robot hardware. The software can easily be adapted to other
quadruped walking machines. An exchange of the Robot module also affects
the Hardware Abstraction layer, which has to be modified according to the
specifications of the robot. However, the MotionGenerator, and thus, the
part of main interest here, can be left unchanged.

3. Flexible statically stable motion

Apart from the flexibility of the software, the goal of the development
process was the construction of a highly flexible walking machine, that
can be utilized independent of the location. Thus, the use of a flexible gait
pattern is necessary. Characteristic for flexible gait pattern is, that the
motion of the platform is governing the motion of the legs. While the plat-
form is moving according to the instantaneous operator's commands, the
instantaneous state of the robot prescribes the choice of the foot that should
perform the next step.3 This can be done following different basic concepts.
One possibility consists in regarding the proximity of the feet to the limits
of the corresponding w ~ r k s p a c e .If~ he,t,f represents an abstracted m e a
surement of the proximity and weot,f the velocity towards the boundary of
the workspace of the corresponding foot f E F,F%et of the four feet, the
foot can be determined by
fstep = {f E FFreeToLift I max(heot,f * Veot,,)} , (1)
where TFfieeToLift is the set of feet, that can be lifted without disturbing
the stability of the robot. For this foot, step length and direction are de-
termined. The step direction is directly prescribed by the operator. In a
first implementation, the step length is a combination of the instantaneous
180

proximity of the foot to the boundary of workspace and the desired velocity:

lstep = kprozbeot + kvelf($, 6,4,z>, (2)

where kprO, and kvel are scaling factors and i, y and GZ

the desired central
body velocities. To ensure a continuous motion of the central body without
deadlocks and to guarantee the possibility to lift the feet, a corrective ve-
locity in opposite direction of the desired velocity and proportional to the
maximal proximity measurement is applied:

ic,,, = -(max(de,,,f . ~,,~,f))~legmove= i, (3)

gcorr = -(ma(beot,f . ueot,f))nlegmowea, 9, (4)
?lzcorr= -(ma(Seot,f . u e o t , f ) ) n l e s m o u e z4 z . (5)
Step length and direction must be transformed in a foot velocity. The
step can be divided into three phases. In the first phase the foot is lifted,
the second phase lasts until the necessary step length is reached, during
the third phase the foot is lowered. Each phase generates velocities to fulfill
the specific function. To receive a smooth foot motion, the three phases
overlap.

4. Quadruped prototypes: ALDURO and ADONIS

The software was initially developed for the quadruped walking machine
ALDURO (see Fig. 2). The robot mainly consists of a 3.5x2.0x2.0m3
(LxWxH) steel structure. With the four legs straightened it reaches a height
of 3.5m. Each leg consists of an upper and a lower part, which are connected
by a rotational joint in the knee. In the hip, each leg has three degrees of
freedom (d.0.f.). The legs are actuated by hydraulic differential cylinders.
The hydraulic system operates at a pressure of 160bar. However, as the
system weights about 1.8t, tests are expansive due to safety measures. To
simplify experiments the small-scale quadruped mechanism ADONIS was
desired. It consists of a 30x40cm2 aluminum structure and four equal legs
as depicted in Fig. 3 and Fig. 4. The main body is holding the electronic
control units, sensors and batteries. It weights about 13kg and is able to
carry an additional payload of 5kg. The main characteristics of ADONIS
are listed in Tab. 1.
In contrast to ALDURO, the d.0.f. of each leg of ADONIS are reduced
to two rotational joints in the hip and one in the knee. The first joint rotates
around a vertical axis that is mounted on the body frame, the other two
 181

Fig. 2. ALDURO

Fig. 3. ADONIS Fig. 4. The leg mechanism of ADONIS

Table 1. Main features of ADONIS

~

LXWXH 1.0x0.5x0.5m~ Total Weight 13kg

Material Aluminium Power Supply 5V DC/24V DC
DC Motor 28Nmm Hip Length 55mm
Planetary Gearbox 4.5Nm Upper Leg Length 220mm
Bevel Gear ratio k2.5 Lower Leg Length 170mm

joints rotate around two parallel horizontal axes. The joints are actuated
by means of DC motors.
Both robots carry a computer similar to a normal PC, a linux system
 182

with embedded real time environment is installed supporting the control

software. ALDUROand ADONIS are both quadruped walking machines.
However, the differences listed in Tab. 2 are significant, which emphasis the
generic character of the control software used for both machines.

ALDURO ADONIS
Dimension LxWxH 3.5x2x3.5m3 1. O X O . ~ X O5m3
.
Total weight 1600kg 13kg
Engine Hydraulic Electric
Degrees of Freedom 16 12
Leg Structure 3 DOF at hip 1 DOF at knee 2 DOF at hip 1 DOF at knee

5. Portability Criterion for Static Gaits

Although both systems differ, the relevant characteristics to use the intro-
duced free-gait algorithm are similar. The choice of a foot to perform a step
depends on its proximity to the boundary of the corresponding workspace.
Thus, for two walking machines using this algorithm, the workspace of the
feet and the distance of hip to the ground should only differ by a scalar
factor. For ALDURO we assume a walking height of about 1.3 - 1.4m,
for ADONIS approximately 0.25 - 0.3m. That leads to a scaling factor of
X M 5. Correspondingly, the dimensions of the workspaces in the walking
planes must differ by the same factor A. As shown in Fig. 5, the shape
of the planes is not surprisingly different due to the different kinematics,
but the scaled workspace plane of ALDURO lays almost completely in the
plane supported by an ADONIS leg. To be able to transfer the experi-
mental resuts gained with ADONIS to ALDURO, the workspace of the
feet of ADONIS is artificially limited by the control software during the
abstraction of the sensor information.

6. Conclusion and Future Work

Quadruped walking machines offer a compromise between fast and flexible,
robust locomotion.
In this paper a generic free-gait algorithm for flexible, statically stable
walking was proposed. This particular algorithm can be used to control
different quadruped machines. It is independent of the actuation type used
for a particular machine and independent of the kinematic structure of
 183

the legs. As a criterion for the portability of the gait pattern, the scaled
workspace of the feet in the walking plane is proposed. The application
of this criterion is shown for the experimental prototypes ALDURO and
ADONIS.
In future research also dynamic gaits for quadrupeds should be ad-
dressed, Here, the question whether portability criterion also for dynamic
gaits exist is of special interest.

Workspace of ADONIS

Workspace of ALDURO

Views of worspaces
in the walking plane

Fig. 5 . Comparison of workspaces for ALDURO and ADONIS

184

References
1. J. Muller, Entwicklung uirtueller Prototypen des hydraulisch angetriebenen
Schreitfahrwerks A L D URO, Dissertation, University of Duisburg, April 2001.
2. D. Germann, M. Hiller, D. Schramm, Design and Control of the Quadruped
Walking Robot ALDURO, ISARC 2005, FERRARA, ITALY, Sept. 11-14,
2005.
3. D. Germann, A Modular Controller Structure for the Quadruped Robot A L -
D URO, Dissertation, University Duisburg-Essen, to be published 2007.
4. S. Song, Y. Chen, A free gait algorithm for quadrupedal walking machines,
Journal of Terramechanics, 28(1), 1991.
 ON THE DESIGN OF A FOUR-BAR MECHANISM FOR
OBSTACLES CLIMBING WHEELS

ANTONIO GONZALEZ
School of Industrial Engineering, University of Castilla-La Mancha, Av. Camilo Jost
Cela s/n, Ciudad Real, 13071, Spain

ERIKA OTTAVIANO, MARC0 CECCARELLI

LARM: Laboratory of Robotics and Mechatronics. DiMSAT, University of Cassino, Via
di Biasio 43, Cassino (Fu, 03043, Italy

This paper describes the design of a mechanism that allows a wheel to climb obstacles
and stairs. The proposed four-bar linkage can be installed on each wheel of a vehicle,
which can be capable to climb stairs with suitable smooth motion. An algorithm is
formulated to design an optimum solution of the mechanism fulfilling task requirements.
A suitable trajectory for the centre of the wheel can be ensured through an easily
controlled motion, and the compactness of the mechanism design makes it suitable for
staircase-climbing wheelchairs for people with disability.

1. Introduction
The number of applications for mobile robots increases continuously. One of the
most interesting and useful application is related to powered wheelchairs. These
vehicles greatly improve the mobility of the people with disability but they may
become useless when face with obstacles, which still exist in many cities and
buildings. In general these obstacles can be modeled as steps, slopes, and stairs,
with different characteristics.
Several systems have been developed to make a mobile robot capable of
climbing and descending steps, slopes and stairs. Conventional wheeled vehicles
have a limited capability for climbing steps and therefore, several solutions can
be used for having efficient and comfortable vehicles. The most used solutions
make use of tracks [ 11 or cluster of wheels [2] or legged systems [3].
Hybrid systems have been also developed as combination of wheeled and
legged solutions to exploit the advantages of both mobile systems. In [4] a leg-
wheel hybrid stair climbing wheelchair is proposed, which consists of eight
independent prismatic-joint legs, four active wheels, and four passive casters. A
prototype reported in [5] consists of a four wheels wheelchair with two legs for

185
186

obstacle climbing.
The CALMOS-Wheelchair prototype (CAstilla-La Mancha Obstacle
Surpassing Wheelchair) has been built at School of Engineering of Castilla-La
Mancha University and it belongs to the category of hybrid locomotion vehicles.
This prototype has been developed to solve drawbacks of existing step-climbing
systems though a new design presented in [7]. The main features of CALMOS-
Wheelchair are high load capability and a considerable adaptability to a wide
variety of obstacle geometries.
In this paper we attach the problem of designing a suitable mechanism, which
can drive wheels with a smooth motion in climbing obstacles and steps. This
paper is focused in the synthesis of a climbing mechanism, whose preliminary
solution has been conceived for CALMOS-Wheelchair prototype. The aim of
the new mechanism proposed in this paper is to increase the adaptability of the
CALMOS-Wheelchair prototype to different obstacle geometries. The system is
also designed in order to be compact and easily controlled.

2. The step-climbing problem for a wheel

The problem of step-climbing can be defined as driving a wheeled vehicle to
climb stairs with suitable smooth motion and efficient power transmission-
mechanisms. The aim can be identified in enhancement of wheelchair
functionality for people with disability. Main characteristics must be preserved
in terms of lightweight design and easy controllability.
It is well known that wheels have limited capacity to climb obstacles. The
maximum obstacle height that a wheel can climb depends on the geometry of
the obstacle and the friction coefficient p. Figure l a shows the trajectory of the
centre of a wheel while climbing a step of height H < r. Figure l b shows the
forces and torques acting on a climbing wheel at the initial configuration of the
climbing process.

a) b)
Figure 1: Wheel climbing a step with height H < r: a) Trajectory of the Wheel centre b) Forces acting
on the wheel.
N is the normal force at the contact wedge and pN is the tangential force. The
 187

weight supported by the wheel is W and the traction torque is T. The needed
friction coefficient p to equilibrate W during the climbing motion can be
obtained from the vertical equilibrium force equations in the wheel. This value
is p = tan(a,,,), where amaxcan be computed as

= COS-’ (I-H/R)
amax

and it is the slope in the initial point of the trajectory point A in Fig. la.
Therefore, the maximum step height H that a wheel of radius r can climb with a
friction coefficient p, can be evaluated by considering Fig. l a as

Similar expressions can be obtained for slopes and H > r steps. The illustrated
situations show that it is necessary to use extra mechanisms when using a wheel
for step climbing. This mechanism will be called Climbing Mechanism.

3. A solution for powered wheels

A Climbing Mechanism must provide a way to maintain the contact of the
wheel with the obstacle and to ensure the needed traction at each configuration
of the climbing process. Therefore, the Climbing Mechanism must fulfill the
conditions: 1) to guarantee traction force for the trajectory without depending on
the friction; 2) to adapt the trajectory of the wheel centre to the obstacle
geometry.
The proposed climbing mechanism in Fig. 2 is composed of a frame attached to
the wheelchair chassis, a four bar linkage, and a sliding support. Friction
coefficients at the wheel contact for usual surfaces in urban environment do not
allow high heights and large slope angles in obstacle-climbing. In order to
overcome this problem, a sliding support can be used, as shown in Fig. 2b). The
so-called sliding support ensures the traction during the obstacle climbing and it
is connected to the chassis by a motorized prismatic joint. The wheel will
remain in contact with the obstacle surface while an additional sliding support is
deployed. This is necessary in order to retransfer the wheelchair weight to the
wheel when the sliding support has terminated its work. Therefore, the wheel
axle cannot be fixed to the chassis. For this reason, an additional DOF must be
considered for the wheel axle to allow its backward movement for avoiding the
obstacle interference and ensuring surface adaptability. Dashed line in Fig. 2b)
shows the relative path of point M as referred to the Climbing Mechanism
frame, and 6 is the slope angle for a desirable trajectory for M.
This trajectory of the wheel centre should be as close as possible to a straight
188

line in order to obtain a smooth and easily controlled movement of the chassis
in the first phase of the climb. The slope of the trajectory must be also set with
descendent, horizontal or ascend direction to ensure the desired behavior of the
mechanism. An ascending trajectory of the wheel has been used for the
CA~MOS-Wheelchair in Fig.2a) to show the operation of the proposed
climbing mechanism. The slope angle of this trajectory is 6 . The main advantage
of this slope is that the climbing or descending process can begin without help
when the added DOF is unlocked. A drawback is that the free barrier wheel
position of the Climbing Mechanism can be movable. Thus, in this
configuration the component of the reaction force in the direction of the
trajectory, may push the wheel to move forward and, therefore the wheel axle
must be locked when the system is in free barriers environments. The four-bar
mechanism is also equipped with a spring to ensure the backward movement of
the wheel to the initial position when the obstacle has been surpassed.
Climbing Mechanism operation in step ascending can be described through the
following phases. The wheelchair stops when the wheel reaches an obstacle and
the sliding support is deployed once the head of the sliding support has reached
the top of the obstacle, the wheel axle DOF is unlocked and the weight is
transferred from the wheel to the sliding support. The sliding support is being
deployed while the added DOF allows the wheel to avoid the interference with
the step. The axle DOF is blocked when wheel goes back to its original position.
Then the sliding support can be folded back. This operation transfers the weight
from the sliding support to the wheel again.

a) b)
Figure 2: CALMOS-Wheelchair with a climbing system: a) a prototype; b) a scheme of the proposed
climbing mechanism.

4. A four-bar mechanism for the ~ l i m b i n g~ e c h a n i s m

 189

A four-bar mechanism has been poposed for the Climbing Mechanism because
of its robustness and trajectory generation ability, besides the characteristics for
lightweight compact design. The proposed four-bar mechanism can be
synthesized to give a desired trajectory of a reference point M of the coupler.
The wheel centre has been chosen as coincident to point M, and its trajectory
must be as close as possible to a straight line with a slight ascending slope 6, as
shown in Fig.2b).
A symmetrical four-bar mechanism has been selected with an approximate
straight-line trajectory. A design scheme is shown in Fig. 3 with b = c = q as to
give a Chebyshev mechanism [8]. Referring to the sizes of the CALMOS-
Wheelchair, Fig. 2, a suitable straight-line segment for M path will be of L =
MM’ = 100 mm. This value allows surpassing the most usual obstacles in urban
environments. The path deviation from a straight-line can be evaluated by the
area denoted by S that can be computed as

where a. and afare the values of input a angle at the initial and final
configurations; x and y are the coordinates of point M with respect to the OXY
frame, being y(a) a general value depending of a and yo is a value at the initial
configuration of the mechanism. When of the origin of the reference frame is
chosen to be coincident with C then yo is coincident with V parameter, and the
X coordinates of center of the segment MM’ is 0, Fig.3.
The compactness of the mechanism is expressed through a function u, which is
the maximum of the distances among points A, B, C, D and M, in the form

u = max (b,CM,CD,MB,MD) (4)

Figure 3: A design scheme for a four-bar linkage whose coupler point M traces an approximated
straight-line trajectory between M and M’.
A maximum range A0 = €)fin - 00 for the swinging motion of the coupler can be
190

also considerer as a constraint to avoid interference among links of the

mechanism.
Two objective functions can be considered for an optimum design of a four-bar
linkage as a climbing mechanism for Climbing Mechanism in CALMOS-
Wheelchair by using the above-mentioned criteria. The main design variables,
are V, b and r. In order to have dimensionless objective functions, S and u can
be suitably divided by S&b and &h&, which are the values of functions S and u
for a Chebyshev mechanism having b/a = 2 and d a = 2.6. Thus, the
optimization design problem for a synthesis of a four bar linkage can be
formulated as
- -

(5)

subjected to

G = Aemm - 2tan-1-L <0

2v
where Kl and K2 are weighting factors that can be conveniently chosen
according to the relationship K, + K2 = 1; A€),,, is the prescribed maximum
swinging rotation of the coupler during its operation, and condition G is
computed by taking into account of the coupler rotation angle as Ll2V.

Table 1. Optimal design results of the optimization design process of Eqs. ( 5 ) and
(6) for different values of the coefficients K,

K1 A0 b r F S U
(rad) (mm) (rad) (mrn’) (mm)
0.9 0.41 247.8 -1.39 0.24 5.4 247.8
0.7 0.58 175.8 -1.39 0.53 16.0 175.8
0.5 1.25 86.2 -0.84 0.51 12.5 117.4
0.3 1.45 75.6 -0.85 0.64 24.8 104.5
0.1 1.57 70.8 -0.86 0.71 35.9 98.5

5. Numerical example
Table 1 shows results of the optimization design process for different values of
Kl, when the maximum rotation of the coupler is limited to 7112. Because of the
low number of design variables a thorough search has been performed in order
to avoid local minima of F. The optimal design solution has been selected with
K, = K2 = 0.5 to have equal consideration of the optimality criteria.
Figure 5 shows the evolution of the objective function F and its components S
 191

and u, together with the design constraint in (6). In the reported example, the
optimization process takes 140 iterations to converge to the optimal solution.
The computation time was 11.23s with a 1.7 GHz Pentium IV using the code
@incon of Matlab Optimization Toolbox [9]. Figure 6 shows a scheme for the
optimal design solution that can be applied to CALMOS-Wheelchair. The
approximate straight line path MM’ is reported with a mechanism simulation
that has been carried out for the layout in Fig. 2b.
5

1
4

0.6 -
3
0.
2

----”.A -*“&L
*
I I 1 t
l t ~ ~ ~ Ite
a) b)
Figure 5: Evolution of the solution for K, = K2 =0.5: a) the objective function F with its components
S and u and constraint G; b) the design variables H, b, r.

b)
Figure 6: Optimal mechanism for the relative motion in figure 2b) for the optimization design
procedure in (5) and (6) for the case with K1= K2= 0.5: a) picture; b) scheme

6. Conclusions
In this paper a Climbing Mechanism for wheelchairs has been proposed with
features of compact design and easy operation. A four bar linkage for the
climbing Mechanism has been designed as attached to each wheel to give
suitable smooth motion during the step climbing. An optimization problem has
192

been formulated to synthesize a proper linkage for the CALMOS-Wheelchair

with satisfactory optimal features in terms of path generation and design sizes

Acknowledgments
The first author wishes to thank the Regional Government of Castilla-La
Mancha (Spain) for the grant that has permitted him to spend a post-doctoral
period at Laboratory of Robotics and Mechatronics in Cassino during 2006.

References
1. K YONEDA, Y OTA, and S HIROSE. Development of a Hi-Grip Stair
Climbing Crawler with Hysteresis Compliant Blocks. Proc. of 4th
International Conference on Climbing and Walking Robots (CLAWAR
2001), pp.569-576 Karlsruhe, 2001
2. M. J. Lawn, T. Ishimatzu. Modelling of a Stair-Climbing Wheelchair
Mechanism with High single-Step Capability. IEEE Transaction on Neural
Systems and Rehabilitation Research. Vol 11. No. 3. pp 323-332. 2003.
3. Giuseppe Carbone, Marco Ceccarelli, Yusuke Sugahara, Hun-ok Lim,
Atsuo Takanishi. Stiffness Experimental Monitoring for WL- 16~ 11Biped
Locomotor During Walking. ROMANSY 16. Proc. of the 16th CISM-
IFToMh4 Symposium on Robot Design. Dynamics and Control. pp. 105-
112. Warsaw, 2006.
4. J. Yuan, S. Hirose. Zero Carrier: A Novel Eight Leg Wheels Hybrid stair
climbing Mobile Vehicle. Journal of Robotics and Mechatronics. Vol. 17.
No.1. pp. 44-51. 2005.
5 . P. Wellman, V. Krovi, V. Kumar and W. Hanvin. Design of a wheelchair
with legs for people with motor disabilities. IEEE Transactions on
Rehabilitation Engineering, Vo1.3. No.4. pp 343-353. 1995.
6. R. Morales, V. Feliu, A. Gonzilez, P. Pintado. Kinematic Model of a New
Staircase Climbing Wheelchair and its Experimental Validation. Int.
Journal of Robotics Research. Vol. 25. No. 9. pp. 825-841.2006
7. A. GonzBlez, R. Morales, V. Feliu, P. Pintado. Improving the Mechanical
Design of a New Staircase Climbing Wheelchair. Proc. of the 9th
International Conference on Climbing and Walking Robots (CLAWAR
2006). pp. 14-18. Brussels. 2006
8. M. Ceccarelli A. Vinciguerra. Approximate four-bar circle-tracing
mechanisms: classical and new synthesis. Mechanism and Machine Theory.
Volume 35. Issue 11, pp. 1579-1599.2000
9. Grace A., "Optimization Toolbox User's Guide", The Matlab Works Inc.,
1992.
 PATH PLANNING FOR THE "3DCLIMBER"

MAHMOUD TAVAKOLI
LINO MARQUES
ANiBAL T. DE ALMEIDA
Institute for Systems and Robotics, University of Coimbra, Portugal
Email: mahmood.tavakoli@gmail.com, [ h o , anibal]@isr.uc.pt

3DCLIMBER is a running project in the University of Coimbra to develop a climbing

robot for climbing and manipulating over 3D human-made structures. This paper mainly
discusses the path planning algorithm of the robot.

1. Introduction
Development of climbing robots was a challenging area during last decade.
Different types of climbing robots were developed for climbing over flat or
curved surfaces. For holding the robot attached to the surface they used suction
cups [ 1,2,3,4,5] or magnets [6,7]. Also robots whose end-effectors match
engineered features of the environment like fences or porous materials or bars
[8,9,10,11,12] were developed. Different kind of robots also developed for
climbing inside pipes or ducts [ 13,141or climbing over poles [ 15,16,17,18].
3DCLIMBER is a running project in the University of Coimbra to develop
a pole climbing robot which is not only able to climb over 3D structures with
bends and branches, but is also able to perform practical and complex tasks like
painting and performing non destructive tests on 3D structures. The robot takes
advantage of a dedicated 4-DOF serial mechanism which enables it to climb
over poles, overcome the bent section and alter between different branches of
the pole. This paper mainly focuses on the path planning and simulation of the
3DCLIMBER, also a summary of the design and kinematic analysis will be
presented.

2. The Robot Design

As M.Tavakoli et. Al. demonstrated before, locomotion and manipulation along
circular 3D structures with bents and branches requires a minimum of 4 DOFs
include 2 translational and 2 rotational DOFs [ 191.
The proposed pole climbing robot consists of two main parts (Fig. 1) which
are: 4-DOF climbing module and gripping module. The robot was designed for

193
194

a step by step movement. One

of the grippers is attached to a
manipulator, and the other one
is attached to the base of a
rotating platform. This
configuration provides four
DOF between grippers,
allowing the movement along
poles with different cross
sections and geometric
configurations. The proposed
design takes advantage of
novelties in the design of all
parts: the climbing module, the
gripping module and the
rotating platform. The climbing
module consists of a 3-DOF
planar Serial arm and a 2; axis
rotating mechanism. Combining
the 3-DOF arm with the
Figure 1. Detailed design of the 3DCLIMBER rotating mechanism provides
two rotations and two
translations, which are necessary to achieve the design objectives as explained
previously.
There are 2 gripping modules. One of them attached to the base of the 4-DOF
climbing module which is indeed the Z axis rotation mechanism and a second
gripper is attached to the manipulator of 3-DOF arm. Each gripper consists of
two unique multi-fingered V-shaped bodies, a motor, one right hand and one left
hand ball screws and 2 linear guides. Table 1 shows the estimated characteristics
of the robot.

3. Kinematics of the Manipulator

Kinematics analysis is necessary in order to perform the workspace analysis,
path planning, optimization and control of the robot.The study of manipulator
kinematics is divided into two parts, inverse (or reverse) kinematics and forward
(or direct) kinematics. The inverse kinematics problem involves mapping a
known pose (position and orientation) of the moving platform of the
manipulator to a set of input joint variables that will achieve that pose. The
forward kinematics problem involves mapping from a known set of input joint
variables to a pose of the moving platform which result from those given inputs.
 195

For performing kinematics analysis, a simplified model of the robot was

considered and Denavit-Hartenberg table was established (fig. 2).
For the direct kinematics analysis, transform matrices were obtained. The
result of the direct kinematics is the transform matrix between the 4'h and the Oth
coordinate systems. Here detailed calculations are not presented but only the
final transformation matrix is presented.
,T=;TX;TX;TX;T
0

O.Scos(V) + 0.5 C,,,, O.Ssin(V) - 0.5 * So Px

=
I:[ O.Ssin(V) - 0.5 Sol,, 0.5 C,,,, - O.Scos(V) C, Py
'123 1' 23 0 pz
0 0 0 1

Letting:
v= -e3-e2+eo-e1
e ~ + coI2
~ ~ = o . ~ ~ ~ ~ c o ~ ( -0.5Lp e ~ +OSL,
- e ~ ) xcos(oo-el)+
+ OSL, x col+
C,XL,
Py, 0.5L2x~in(-02+00-81)+
0.5L2x SOl2+0.5L1xsin(00-01)+0.5L1x Sol+SoxLo
Pz = SI2XL2+ SlXLl
For inverse kinematics problem the position matrix [Px PY P z ] ~of the 4th
coordinate system and also 0 which is the angle of X4 and XI, around Z4, were
considered to be known. The values of 0, to €I3 should be obtained. For a known
pose 2 different set of answers obtained, while for both of them 0, is same.

r- 4 TABLE II
MAIN CHARACTERISTICS OF THE DETAILED

Degrees of Freedom 4
. , I
Step by
Climbing Procedure 9t-n

Approximate Weight /kg) I 40 -1

Robot Size (cm) I 50*60*50 I
Extended Robot Size (cm) 50*60*100
Theoretical Climbing
Speed (dmin)

Figure.2. Coordinate systems

196

e0 = A tan 2 5 , B Z l = A tan 2(+,/=,Q),e2, = A tan 2 ( - , / 3 , Q )

PX
el, = A t a n 2 ( P z , C , P x + S o P y - L , ) - A t a n 2 ( k 2 , , k , , )
el, = A tan 2 ( P z , C 0 P x + SOPy- L o )- A tan 2(k2,, k , , )
831 = e - el, - e,, - e3,,e3,= e - el, - e,, - e,,

Letting:

o=Px2 +Py2+Pz2+Lo2-L12- L I Z -2L,.\IPx2 +Py2

Two functions were established in MATLAB, to calculate inverse and

direct kinematics. These functions will be used in all further steps like
workspace analysis and path-planning. Validity of equations and functions was
checked by numerical examples. Constant orientation workspace analysis also
performed by numerical approaches in MATLAB. The results were specially
used for determining the appropriate links lengths of the 3-DOFs arm which
make it possible to overcome the bent section with angles up to 100 degrees.

4. Path Planning and Simulation

Selection of an appropriate Path planner is a critical issue. In Robotics
application, path planning and path tracking can be done online or offline
(Figure 3). This depends on several factors include existence of obstacles,
predictability level of obstacle locations and the amount of knowledge available
from the robot working area before start of the robot operation [20].We may
plan paths before their execution, if we have some knowledge of the
environment. Planning paths before execution allows efforts to get a shorter
path time, more efficient dynamics, and absolute collision avoidance. When
working in this mode a priori knowledge (i.e. Known before) is used. Priori
knowledge may not be used for unpredictable or random motion as there is no
detection method allowed by its definition.
For our application we should decide for using a Priori path planner or a Postieri
path planner or combination of them. As stated earlier the working area of the
3DCLIMBER robot is human made 3D structures. So the geometry of the
structure and thus the working area of the robot are known beforehand. Also
proposed tasks of the robot are climbing and manipulating over structure. All of
these tasks are performable with composition of 3 more basic tasks.
 197

A. Climbing over the straight part of the structure. It can be done with
straight line path planning algorithms.
B. Rotating around the axis of the structure. It is necessary for scanning all
points of the structures, as some of the applications need.
C. Passing the bent section of the pole, and performing some complicated
tasks like welding need rotation of the manipulator around Y axis.
Regarding these 3 parts and descriptions, it is possible to plan the movement of
the robot in priori mode. But we should always consider errors during the
operation. The most significant error is slippage of the robot. Also as the precise
geometry of the structure may not bee known beforehand, it should be measured
during operation with sensors and in this case a complete priori path planning is
not possible because of errors and lack of the precise information about
structure. To overcome this problem a higher level Priori planer should be
combined with a lower level Postieri planner. This combination is called
Hierarchical planner. It means to design an offline path planning plus an online
path tracking (Figure 4). For instance when the robot is climbing along the
straight part of the structure, the upper level Priori planner send the signal for
the straight path planner module and this module send appropriate signals to all
motors of the robot for going one step forward. This module also includes an
online path tracking which compensate available errors.
Collision
Avoidance
Yes No
Off-line Collision free Off-line Path
path planning plus, cm- Planning, plus on-line
Yes line path tracking. path tracking
Path
Constraints
Positional Control. Positional Control,
plus on-line obstacle
No detection and
avoidance.

Figure. 3. Chart for selection of path planning method

Low Level Priori and Postieri Planners

Straight path Planner Module:

-Q- Bent section Planner Module

Pole AxjS: rotation PIShrter Module:
Test,and Maintenance'Planner,Modules

Figure 4. 3DCLIMBER path planner chart.

198

When the robot reaches to a specified distance from the bent section of the
pole, the higher level planner receives information from the sensors about the
exact distance of the manipulator from the bent section and also the angle of the
bent section. Then it sends this information to the bent passing module. This
module should be a Hierarchical planner. It should receive the information of
geometry of the pole and plan the control of the motors during the operation of
the robot.

4.1. Straight Path Planner Module

The climbing strategy is opening of the top gripper, moving up the manipulator,
closing of the top gripper, opening of the bottom gripper, moving up the base
and closing of the bottom gripper. For the robot to travel along straight path, a
trajectory planner function was developed in MATLAB. This function generates
appropriate data for control of motors in the 4DOF mechanism and grippers so
that the robot travels along a straight path. Then this function was used for
simulation of the robots model which was developed in Solidworks. Cosmos
motion was used as simulation package.
Priori planner generates necessary data as input for this module. This data
include start point and end point of one climbing step in the Cartesian order.
The output of the module is the matrix of the position versus time for the 4DOF
mechanism's motors, and torque versus time for the gripper's motors. The
applied algorithm estimate n points on the straight path and by applying inverse
kinematics it generates appropriate data for all of motors.

4.2. Bent Section Passing Module

An algorithm also developed for passing the bent section. This algorithm
receives the distance of the manipulator from bent section and bent angle from
world modeler and generates data for robot joints for passing the bent section
without collision.

4.3. Simulation Results

The geometry of the pole and the start position of the robot were given as input
to the path planner and the final position of the manipulator for the maximum
possible thrust was calculated. Using the developed functions the database of
position versus times was generated for all motors of the 4 DOF mechanism,
and also database of torque versus time was generated for gripper motors. Then
this data was given as the input data to the cosmos motion simulation package
and simulation was started. After each step the new position of the manipulator
and base of the robot was reported to the main algorithm. In this simulation
process there is no difference between the desired and the resulted position, but
 199

in the real robot there would be always an error due to different sources of
error and mainly slippage of the grippers. To solve this problem in the real
robot laser triangulation sensors will be used in the base and manipulator of the
robot and after finishing each step, the resulted error would be reported to the
tracking algorithm and would be compensated. Figure 5 shows generated data
from developed hnction for passing the bent section. X, Y and Z are position of
the manipulator and €3 is the angle of manipulator in respect with base of the
robot. Oo to O3 are angles of the 4 DOF mechanism, which are showed in both
abstract and relative mode. Also curves of generated data for gripper exerting
force versus time are showed. Figures 6 show some pictures from the simulation
of the robot.

0 5 10 15 20 25 31
Time (sec) Time (sec)
I
a
........:*Gripper ofthe base

e 0 5 10 15 20 25 30
Time (sec) Time (sec)

Figure.5. Generated data for passing the bent section

5. Detailed Design
A pole climbing robot was designed, based on the designed 4-DOFs arms and
the obtained results from simulation. The proposed design takes advantage of
novelties in the design of all parts: the climbing module, the gripping module
and the rotating platform.Each gripper consists of two unique multi-fingered V-
shaped bodies, a brushless motor, one right hand and one left hand ball screws
and 2 linear guides. V shaped grippers has the mechanically self centering
advantage. With other kinds of grippers, there would be a need for sensors and
200

Figure 6 . Simulation of the SDCLIMBER for thrusting along pole and passing over the bent
control strategies to precisely control the position of robot, in order to perform
safe gripping. Weight optimization was considered in the design of all of the
robot’s parts as the weight is very important factor in climbing robots. Harmonic
Drive gearboxes were used in the 3-DOFs serial arm in order to reduce the
weight and increase the precision and the efficiency. All non standard parts
were designed and manufactured with 7075-T6 aluminum. Both grippers and
rotating mechanism were assembled and tested. The 3DOF mechanism is still
under tests. Figure 7 shows the grippers and rotating mechanism of the robot.
 201

6. CONCL~S~ON
3DCLIMBER is a project for
developing a robot able to climb and
manipulate over 3D human-made
structures. In this paper a short
susnmary of the conceptual design of
the robot was described. Then the
path planning algorithm and the
simulation results for the
3DCLIMBER model was presented.
Future works include, multi-criteria
optimization with genetic algorithms,
assembly, control and test of the
robot.

Fig. 7. Grippers and rotating mechanism of the

robot.

1. H. Dulimarta and R. L. T m a l a , Design and control of miniature climbing

robots with nonholonomic constraints. In World Congress on Intelligent
Control and Automation, Shanghai, P.R.China, Jun 2002.
2. A. Nagakubo and S. Hirose. Walking and running of the quadruped wall
climbing robot. In IEEE Int. Conf. on Rob. And Aut., volume 2, pages
1005-1 0 12, 1994.
3. M. Rachkov. Control of climbing robot for rough surfaces. In Int.
Workshop on Robot Motion and Control, pages 101-105,2002.
4. S. W. Ryu, J. J. Park, S. M. Ryew, and H. R. Choi. Self-contained wall-
climbing robot with closed link mechanism. In IEEElRSJ Int. Conf. on
Int. Rob. And Sys, Maui, HI, 2001.
5. W. Yan, L. Shuliang, X. Dianguo, Z. Yanzheng, S. Hao, and G . Xuesban.
Development and application of wall-climbing robots. In IEEE Int. Conf.
on Rob. And Aut. Detroit, MI, 1999.
6. J. C. Grieco, M. Prieto, M. Armada, and P. G . de Santos. A six-legged
climbing robot for high payloads. In IEEE Int. Conf. on Cont. App.,
Trieste, Italy, 1998.
7. S. Hirose, A. Nagabuko, and R. Toyama. Machine that can walk and climb
on floors, walls, and ceilings. In CAR, pages 753-758, Pisa, Italy, 1991.
8. D. Bevly, S. Dubowsky and C. Mavroidis. A simplified Cartesian-computed
torque controller for highly geared systems and its application to an
202

experimental climbing robot. ASME J. of Dynamic Systems,

Measurement, and Control, 122(1):27-32,2000.
9. Y. Xu, H. Brown, M. Friendman, and T. Kanade. Control system of the
selfmobile space manipulator. IEEE Tr. on Cont. Sys. Tech., 2(3):207-
219, 1994.
10. M. Yim, S. Homans, and K. Roufas. Climbing with snake-robots. In IFAC
Workshop on Mobile Robot Technology, Jejudo, Korea, 2001.
11. H. Amano, K. Osuka, and T.-J. Tam. Development of vertically moving
robot with gripping handrails for fire fighting. In IEEElRSJ Int. Conf. on
Int. Rob. And Sys. Maui, HI, 2001.
12. C. Balaguer, A. Gim’enez, J. Pastor, V. Padr‘on, and M. Abderrahim. A
climbing autonomous robot for inspection applications in 3d complex
environments. Robotica, 18:287-297,2000.
13. W. Neubauer. A spider-like robot that climbs vertically in ducts or pipes. In
Int. Conf. on Int. Rob. and Sys., pages 1178-1185, Munich, Germany,
1994.
14. T. RoDmann and F. Pfeiffer. Control of an eight legged pipe crawling robot.
In Int. Symp. on Experimental Robotics, pages 353-346,1997.
15. M. Almonacid, R. Saltar‘en, R. Aracil, and 0. Reinoso. Motion planning of
a climbing parallel robot. IEEE Tr. on Rob. and Aut., 19(3):485489,
2003.
16. Z. Ripin, T. B. Soon, A. Abdullah, and Z. Samad. Development of a low-
cost modular pole climbing robot. In TENCON, volume 1, pages 196-200,
Kuala Lumpur, Malaysia, 2000.
17. M. Tavakoli, M.R. Zakerzadeh, G.R. Vossoughi, S. Bagheri, “A hybrid
Pole Climbing and Manipulating Robot with Minimum DOFs for
Construction and Service Applications”, Journal of Industrial Robot, March
2005.
18. Ali Baghani, Majid Nili Ahmadabadi, Ahad Harati, “Kinematics Modelling
of a Wheel-Based Pole Climbing Robot (UT-PCR).” 2005 IEEE
International Conference on Robotics and Automation, April 18-22
Barcelona.
19. M. Tavakoli, M.R. Zakerzadeh, G.R. Vossoughi, S . Bagheri, H.Salarieh, “A
Novel SeriaYParallel Pole ClimbingDvlanipulating Robot: Design,
Kinematic Analysis and Workspace Optimization with Genetic Algorithm”,
2 1st International Symposium on Automation and Robotics in
Construction, 21 to 25 September 2004, Jeju island, Korea.
20. K.S.Fu, R.C.Gonzalez, C.S.G.Lee, “Robotics; Control, Sensing, Vision, and
Intelligence”, NewYork: McGraw Hill, 1987.
 ROBOT FOR MOTION IN TUBE*

JATSUN SERGEY
Mechatronical Department, Kursk State Technical University,
50 let Oktyabrya, 94,
Kursk, 305000, Russia

MISHENKO VLADIMIR
Mechatronical Department, Kursk State Technical University,
50 let Oktyabrya, 94,
Kursk, 305000, Russia

JATSUN ANDREY
Mechatronical Department, Kursk State Technical University,
50 let Oktyabrya, 94,
Kursk, 305000, Russia

Mobile robots are widely utilized for various operations in environments inaccessible to a
human or dangerous for him. They are utilized, for example, for inspection and repair
work in nuclear power and chemical plants, operations in areas of wreckage after
earthquakes or blasts, or dismantling explosive devices. Most of these robots move by
means of wheels or caterpillars, some of them utilize walking mechanisms. Such robots,
however, cannot enter narrow slots (for example, during rescue operations in a zone of
wreckage) or move in dense media other than gases or liquids. This justifies looking for
new concepts of motion to enable robots to move efficiently in environments inaccessible
to robots with wheel, caterpillar, and walking propelling systems. This issue is especially
topical for medical robots designed for the motion through rather narrow channels (e.g.,
in blood vessels or the intestines) or among muscles to reach an affected organ to
perform a diagnostic or surgical operation.

1. Introduction
In the present paper, the concept of vibration-driven robots is developed. Such
robots can move in various media without wheels, caterpillars or legs. The

*
N205-08-33382and DFG project NcZi. 540/6-
Work supported by RBRF project N904-01-04002,
1.

203
204

propulsion of the robot is provided from one hand due to vibration of internal
masses inside the robot and from another hand due to interaction of the robot’s
body with the environment.
The vibration-driven robot is designed as a mechatronical system consisting
of the mechanical, electrical, and electronic (microcomputer) components. The
mechanical vibration is transmitted from the vibration exciter to the robot’s
body, which interacts with the environment with some force. The robot is
equipped with a feedback microcontroller to maintain an efficient operation
mode and provide prescribed characteristics of motion of the end-effector under
the action of various disturbances. The characteristics of the exciter vibration
should be tuned to a specific task to be executed by the robot. This tuning can be
provided on the basis of parametric optimization in which an operating
characteristic of the robot (e.g., the average velocity of the robot) is used as the
objective function.
Classification of vibration - driven robots is proposed with point of view of
dimension of space where the robot and his parts moves. In simplest case the
robot moves in one- dimensional space (1-D). In this case motion of the robot
can be provided with help of variable shape of body (this is worm like motion
[l-6, 111) or with motion of internal vibrating masses. More complicate case is
the motion of robot in two - dimensional space (2-D). The motion of the robot
can get with variable shape of body with open structure or close structure (snake-
like motion [9]) or with mobile internal masses that can move with accordance of
planar trajectory (very interesting case -is the rotational motion of masses [12]).
The most complicate is motion of robot in three -dimensional (3-D) space. As it
was in previous cases the motion of robot body can get with periodically variable
shape of body with open or close structure or with mobile internal masses.
2. Mathematical model and design of 1-d robot with variable shape of robot
body.

The 1-D motion of the robot can be provided by periodical relative motion of
two parts of robot body along the line (see Fig. 8). In this case, it is necessary
that the characteristic of friction force between robot body and the supporting
surface should be asymmetric. The robot under consideration was designed for
cleaning pipelines from solid deposits.
To determine the parameters of the cleaner robot moving in tube, we
developed a mathematical model that describes the interaction of the end-
effector with the technological load (due to the resistance of the solid deposits)
and the vibration-driven actuator. This robot has control system that provides
 205

adequate electrical feeding of each drive (EM). Control system consists of sensor
(S) microcontroller source (MC) of electrical feeding (PS).

Fig. 1 Calculating scheme of mini-robot with two asymmetric friction elements.

XI ,X2 - generalized coordinates of masses; C and p - spring rate and damping coefficient; C2 , p -
spring rate and damping coefficient of limiter; Q I ,-force
~ generated by the electromagnetic exciter;
P1,2-force of elastic element; R1.2 - force of resistance; AX - length of limiter; A distance between
masses without reference of spring deformation.

Differential equations that describe dynamics of the robot with an

electromagnetic vibration actuator in not dimensional form can be introduced in
a form in Eq. (1):

Where are:
T T

g=-m*
P
m,.o

Where are: o - frequency of electromagnetic drive, Uo, (Do, Xo - scales of

electrical voltage, magnetic flux and length, @ is the magnetic flux; I - is the
current,, R- is the active electric resistance of the solenoid;u(t) is the power
supply voltage, S-square of the air gap inside if electromagnet; -
electromagnetic constant.
206

The system of nonlinear differential equations describing the motion of the

robot is solved by numerical methods. Based on the simulation results, we
determined the parameters of the vibration exciter and analyzed the motion of
the robot values of the friction coefficient.

Fig. 2. The average velocity of the robot as a function of the relative frequency of the vibration
exciter for different electrical feeding. 1-~=6,2-&=4,3-~=3,4-~=2.

Fig. 3. Time history of velocity of robot masses for E =2 and c=0,6,

1 - velocity of first mass, 2 -velocity of second mass.

Figure 3 shows the behavior of robot for different regimes of motion.

Analysis of these results shows that in a field of first resonance masses move to
each other and each masse has stop regimes in every step with frequency that is
equal to external frequency. In a field of second resonance the masses of robot
move to each other and each masse has stop regimes in every step with
frequency less than external frequency in two times. And in a field of third
resonance the masses of robot move to each other and each masse has stop
regimes in every step with frequency less than external frequency in three times.
The some results of the calculation of the force of friction and the average
velocity of the robot are illustrated in fig. 3. Analysis shows that the value of
electrical voltage plays very important role for average velocity of the cleaner
robot. Besides is the ratio of the natural vibration frequency of the robot to the
excitation frequency of the electromagnet is very important also.
 207

3. ~escriptionof 1-d robot with two masses and electromechanicaldrives

This robot consists two parts as it is shown on fig.4. The parts of the robot have
three special friction elements 2 providing asymmetric friction forces. Electrical
coil 1 switches direction of friction element. Electrical drive 3 displaces one part
relatively another part. Due to asymmetrical friction forces acting between robot
parts and tube robot moves inside in pipe. Friction element using in the robot has
boll and spring. The boll moves on the inclined surface. The angle between this
surface and tube surface is smaller than friction angel. It provides asymmetric of
friction force dependence on direction of motion of robot body. For example, the
robot, which introduced on fig. 4 moves to the left side. But if is needed to
reverse direction of motion of robot control system by use of electrical coil will
change inclination of the supporting surface of the bolls for each friction
elements and robot will move to the right direction.

I
I
I ir - robot

I
I I
tube
Fig. 4. Scheme of robot with two masses and electrical drives inside in tube.

Fig. 5. The 3-D scheme ofrobot with two masses and electrical drives inside in tube.

Consider 3-D scheme of vibration-driven robot with two masses that move
and change the shape of the body and generate friction forces is shown on fig.5.
Friction elements have control system. This makes it possible to control
component of the force exerted on the robot by the supporting surface. And it
208

provides reverse motion of the robot inside in tube. In the robot three separate
electrical drives is used. It is necessary in the case when robot moves inside in
curve tube.
In mechatronical department of Kursk State Technical University the
prototype of robot for in pipe inspection is made (see fig.6). This robot moves in
tube with constant internal diameter only.

Fig.6. The general view of the robot with two masses and electrical drives inside in the glass tube
(experimental device).
Control system of vibrating 1-D robot should decide different tasks main of
them could be introduced in the next form:l) Trajectory planning; 2) Control
velocity of motion of robot with accordance of given trajectory; 3)Opti~zation
of velocity of robot; 4) Control of DC- motors angular velocity and phase shift
angle.
In this paper control system for the I-D robot with two masses is introduce
in fig.7.
 209

I I
Fig. 7. Scheme of control system of the robot

Control system consists from microcontroller Atmega8, sensors, connectors

LB 1649 for DC-motors, remote control with infrared sensor. With accordance of
signals from navigation system this control system provides change of velocity
of robot.

4. Conclusions
Vibration-driven robots with a one-coordinate vibration exciter can move along
a rough surface only if the friction between the robot body and the supporting
surface or/and the exciter vibrations have asymmetric characteristics. Robots
equipped with vibration exciters and internal masse can move even in the
absence of an asymmetry in the friction or vibration characteristics. .
Robots equipped with a one-coordinate exciter and two masses has special
friction elements providing asymmetric friction between robot body and internal
surface of the tube and it can be utilized for cleaning pipelines from solid
deposits in a case of resistance is high. The reverse in the direction of motion is
provided by changing of direction of friction forces providing by friction
elements.

Acknowledgements

This investigation is supported by RBRF project NeO4-01-04002, NeO5-08-33382

and DFG project NeZi. 540/6-1.
210

References

1. Aoshima, S.; Tsujimura, T.; Yabuta, T.: A miniature mobile robot using
piezo vibration for mobility in a thin tube, Journal of Dynamic Systems,
Measurement, and Control Vol. 115 (1993), pp. 270-278
2. Ma, J.; Lo, M.; Bao, Z.; Wang, A.: Micro peristaltic robot simulating
earthworm and its control system, Journal of Shanghai Jiao tong University
Vol. 33 No. 7, (1999)
3. Jatsun S., Safarov J. Vibrating engine for robots. Proceedings. CLAWAR
(2000). Madrid.P.1016-1021.
4. Jatsun S., Jatsun S., Vorontsov R. Dynamics of vibrating robot for in pipe
inspection International Symposium - SYROM. - Bucharest, (200 1). 205-
209.
5. Zimmermann K., Zeidis I., Steigenberger J., and Huang Jianjun. An
approach to the modelling of worm-like motion systems with afinite number
of degrees offieedom, First Steps in Technical Realization. Proc. of the 4th
International Conference of Climbing and Walking Robots, Karlsruhe.
(2001). P. 561- 568.
6. Yeh, R.; Hollar, S.; Pister, K.S.J.: Design of low-power silicon articulated
microrobots, Journal of Micromechatronics, Vol. 1, Num. 3, (2001). P. 191-
203
7. Bolotnik N.N., Chernousko F.L., Kostin G.V., and Pfeiffer F. Regular
motion of a tube-crawling robot in a curved tube, Mechanics of Structures
and Machines. (2002). Vol. 30. No. 4. P. 431-462.
8. Chernousko F.L. Snake-like locomotions of multilink mechanisms, Journal
of Vibration and Control. (2003). Vol. 9. No. 1-2. P. 235-256.
9. Zimmermann K., Zeidis I., and Pivovarov M. Dynamics of a nonlinear
oscillator in consideration of non-symmetric Coulomb dry Fiction, Fifth
Euromech. Nonlinear Dynamics Conference. Book of Abstracts. Eindhoven
Netherlands, August 7 12. (2005). P. 308
10. Zimmermann K., Zeidis I., and Steigenberger J. Mathematical model of
worm-like motion systems with finite and infinite numbers of degrees of
fieedom, Theory and Practice of Robots and Manipulators. Proceedings of
the 14th CISM IFToMM Symposium (RoManSy 14). (2002). P. 07 16.
11. Chernousko F., Zimmermann K., Bolotnik N., Yatsun S, Zeidis I. Vibration
-Driven Robots, The Workshop on Adaptive and Intelligent Robots: Present
and Future. Proceedings. Vol. 1 The Institute for problem in mechanics RAS.
MOSCOW. (2005), P.26-3 1.
12. Bolotnik N. N., Jatsun S. F., Jatsun A. S., Cherepanov A. A. Automatically
controlled vibration-driven robots, Proceedings. International Conference
on Mechatronics ICM, Budapest, (2006), P.438-441.
 ROBOTRAIN AS SNAKELIKE ROBOTIC SYSTEM
WITH MINIMAL NUMBER OF DOF

V.E.PAVLOVSKY, V.V.PAVLOVSKY, Jr
Keldysh Institute ofApplied Mathematics ofRAS
Moscow, Russia, 12.5047, Miusskaya sq., 4

N.V.PETROVSKAYA, V.V.EVGRAFOV
Moscow State University,
mathematics and mechanics department, Moscow, Vorob'evi Gori

The paper deals with investigating of dynamical properties and elaborating of

corresponding motion planning algorithms for wheeled "robotrain", which is snakelike
transport structure. In this structure there are one active robot, equipped with differential
drive, and large number of trailers. The created control algorithms provide purposeful
motion of the system along the given set of lines, or inside comdors, or via target points.
Algorithms are verified by means of mathematical modeling. The work is supported by
the RFBR grants No 02-01-00750,04-01-00065.

1. Introduction
The problem of synthesis of control algorithms for snakelike systems of "the
chain of robots" type is concerned with the need which has appeared now in
view of creation of automatic multiobject transport systems. For example, such
robotized systems are good for transportation and delivery of large number of
scientific devices in zones where presence of the human is impossible or is
difficult, and in similar cases. The great value has also creation of robots-chains
as usual transport systems. Note, similar problems arise, in particular, at
robotized service operations in extreme environments. So the problem of control
of snakelike robots motion as chains of mobile objects is an important one and
has theoretical and applied value, researches in this area are stimulated by the
numerous applied problems.

In the paper there are proposed and verified by means of computer simulation
the model and control algorithms for "robotrain", which is a system of one active
robot and a chain of trailers. The algorithms provide purposeful movement of
such chain of mobile objects along the target points set and, so forth, in the

21 1
212

environment with obstacles (with their avoidance). The paper proposes a set of
algorithms, the obtained results can be used on development of control
algorithms for real transport systems with the tractor-robot and the chain of
trailers providing solution of spectrum of applied problems. Note, in the system
tractor-robot may be located not only as the first robot of a chain, but may be set
in any position in a chain, even may be the rear object. The active robot is one
equipped with differential drive and the whole chain have only two DOF, which
is really minimal number of controlled parameters for snakelike system.

2. Robotrain model. The equations of motion, the solutions.

Let's note, that works in this direction actively started over the world last years.
The analysis of the specified works [ 1-4,8] show, nevertheless, that practically
all system have a large numbers of DOF, typical solution is to use drives in each
joint or in each object, e.g. KORYU-I1 is such system. In this sense those
systems are redundant in view of number of DOF, and the question is a
following - what minimal DOF quantity is sufficient for organizing the snakelike
robot motion. In the paper there will be below proposed a snakelike system with
only two DOF, and it is possible to show that those system is quite enough to
built any purposeful snakelike motion on a plane. It seems, this result is an
original one.

Let's consider the snakelike robotic system "robotrain", namely, a robot with set
of trailers. Let's consider following mathematical model of that chain of mobile
wheel objects (carriages), see fig.1. Each carriage contains a weightless axle of
length 2 with wheels of weight m and radius r, a weightless directing rod of
length 2b, on perpendicular axe, and a heavy body of mass mi. The center of
mass of the carriage coincides with the center of an axle of wheels. Carriages are
coupled with each other by means of the cylindrical hinge, gap in coupling is
absent. Carriages move on absolutely rough horizontal plane. The condition of
rolling without slippage causes presence of nonholonomic constrains.

Let 8, 82.., 8 be angles of turn of directing rods of corresponding carriages

around of a vertical axe (angles of orientation), in the further instead for angles
8, ..., 8, we shall consider angles of turn of a directing rod of i-th carriage
( i = 1,2, ...,n ) concerning a directing rod of the previous carriage l,!f2 ,...,l,!f,,,
where l,!f2 = 8, - 8, ,...,W, =8, -en-, , let X I , Y I , ..., Xnr yn - are
coordinates of points CI, ..., n accordingly, qIlq12, ... v,,,, vnz-angles of
 213

turn of wheels of corresponding carriages around their axles.

Figure 1. Projection of robotrain system onto a horizontal plane.

The equations of movement of such system with n elements are made in the form
of Voronec nonholonomic equations, and in the form of Cauchy for free
movement have a following appearance:
%I=@,I

rpIZ = @I2

...
214

( 2 a b c o s i ( - l ~ y l k+(a2-bz)sini(-lyylk)Glz+(a' + b z ) s i n i ( - I y y k
k=Z k=2
G,,
r k =2 ,i=2p
q. = 2ab
-(a2 + bz)sini(-lYyk GI,+ (2abcosA(-Iyylk - (a2- b z ) s i n i ( - l y y k ) G l l
r k=2 k=2 k=Z ' , i = 2 +1
2ab

-(u2 + b 2 ) s i n ~ ( - 1 ) ' y , $ , , + ( 2 a b c o s ~ ( - l-(u2

~ y k- b ' ) s i n ~ ( - l ) * y , ) $ , ,
k=2 k=2
r k=Z ,i=2p
2ab
Vp12 ='
( 2 a b c o s ~ ( - l ) ' y k +(a2- b z ) s i n ~ ( - l ~ q f k ) $ 1 2 + ( u+2b 2 ) s i n ~ ( - 1 ~ y k r p ' , ,
k=2 k=Z k=2
r ,i = 2 p + 1
2ab

Just from the equations (1) it may be obtained important conclusion, that change
of number of objects in system does not influence on change of number of
degrees of freedom which is always equal to two for any n elements.

Figure 2. Examples of trajectories of ballistic movements.

Base solutions of the equations of motion (1) are obtained (see [5-71). It is
necessary to note, that such partial solutions as "straight line'' and ''circle" exist
for any n. Also there exist dynamically smooth solution as Cornu spiral [5-71.
Other possible solutions are found by numerical methods, some their examples
are resulted on fig.2.

Let's consider operated movement of a chain of the mobile system. We shall

consider in this case, that there is only one active carriage (an active robot) and
its wheels are set in motion by engines, and the engine is characterized by the
torques developed by them (Mom,and MomZ).Then the dynamic equations will
become:
 215

.. =
(C,Moq - 2C, MU&) (C,K,-2C3 KI)
+
(C,K,-2C3 K',) ' ' (C, K6-2C3 K , ) "
91'
(2)
(C;-4C, CJ- c;4c,c3 c'2+ c;4c,c3 9'2"'+ c;-4c,c,

where C', ,C', ,C, ,K', ,K', ,K', ,K', ,K', ,K', - known hnctions of v . ..v,.

Using the received equations (1)-(2), we shall construct the alphabet of base
movements of robotrain. Into a class of considered program movements enter: a
straight line, a circle and a Cornu spiral [7]. It is possible to prove that this class
is sufficient for constructing any dynamically smooth trajectodes of robotrain.
Movement on a straight line is set by a condition of equality of angular speeds of
the leading carriage:
. .
(plI = (p,* = const = @

Movement on a circle is set by a following relation for angular speeds of rotation

of wheels of the leading carriage:
. vI2
rp,, =krp,,,
. =consf=$
The linear law of change of speeds of rotation of wheels of the active carriage
CPI, = k , t + w l l > I P I * = k Z f + ~ 1 2

sets movement on a Cornu spiral [5-71.

In a considered class of movements it is possible to obtain laws of change of
torques on active wheels which satisfy dynamic equation (2) - on the basis of the
assumptions made above the operating torques are defined fiom the equations
(1) and (2) as known functions of time.

3. Method for planning of robotrain motion

For further trajectory design we will introduce new base trajectory - a double
Comu spiral which is a pair of base spirals joined in some point and ratio k2 / k,
will denote ratio of lengths of those spirals. Using the base trajectories (a straight
line, a circle, a double Comu spiral), we shall construct methods of planning of
movement of a chain "robotrain" at travel between various points [7]. It appears,
that realization of dynamically realizable transition from one point of phase
space to another by means of unique program movement from class described
above is impossible. The opportunity of such transition is provided at the
combined movement, for example on two spirals.
216

The central technique is planning a trajectory as a set of pieces of a straight lines

with dynamically smooth coupling. And in general it is necessary to plan such
movement only for active robot, motion of the rest part of system will occur in a
vicinity of that set of pieces. The deviation of the rest part of chain from the
trajectory of active robot is precisely known from (1)-(2) and may be also taken
into account by appropriate choice values of free parameters, as shown below.
Note, pre-given set of pieces of lines may be produced, e.g., by geometrical
route planning system, etc. The type of a similar trajectory is shown on fig.3. In
fact just the stages of construction of resulting trajectory for active robot are
shown on fig.3.

Let I, 11, 111 and IV on fig.8 are pre-set pieces of given trajectory and let's denote
them as 0102, 0203, 030, 0405correspondingly. In particular, movement on a
trajectory, which is set by two pieces I, 11, is broken into three sites:
- movement on a straight line setting the first piece 0102, from an index point
O1up to a point Al which is setting on independently on the first straight line;
- movement along two arcs of spirals (on a double Cornu spiral) from a point Al
up to a point A2 belongs to a straight line setting the second piece of movement
0203, point C of coupling of spirals is defined by given ratio k2 / kl ;
- movement on a straight line setting the second piece, from A2 up to the target
point 03.
4'

Figure 3. Stages of planning of robot movement.

And so forth for smoothing the coupling of pieces 0203, 0304, 0 4 0 ,see fig.3.
Evidently, the resulting trajectory will satisfy as for dynamic equations of
motion, as for equations of constrains (1)-(2). So, the resulting trajectory will be
dynamically realizable. It is also necessary to note, that this technique has the
following free parameters: positions of the points A,, A2,and C (i.e. the ratio k2 /
kl ), with using those parameters it is possible to built optimal smoothing of
trajectory pieces, for example, in the sense of values of deviation resulting
trajectory from given set of pre-planned lines. This means, it is possible to
optimize accuracy of resulting trajectory, and so on.
 217

With use of this feature two principles of planning of trajectory of considered

system were developed: a principle of "corridors" and a principle of "reference
points". First one takes into account the permissible values of deviation of
system from pre-set lines. In case of a principle of "reference points" there is
given a numbered set of points on a plane and the active robot has to go via all
that points. The orientation of active robot is given in each point as well. It is
easy to show then by constructing an auxiliary strait trajectories this problem
may be reduced to a problem of planning of movement by a principle of
"comdors" .

Figure 4. Examples of the planned trajectories.

Results of work of algorithms of planning of robotrain movement are shown on

fig.4. The trajectory in the right figure should pass precisely through the given
points, it has greater average curvatures.

To verify the results discussed above the another model was elaborated. That
model was developed as multibody dynamics model, an example is shown on a
figure below.

Figure 5. Simulation samples.

It is necessary to note, that simulation with using multibody model is more

accurate because allow to take into account several additional details, such as
more realistic contact between wheel and a surface. In fact, this model also show
good functioning of the elaborated system.
218

4. Conclusion
Numerical experiments show good quality of elaborated control methods. From
those results it is possible to conclude that proposed techniques of planning and
realizing robotrain motion are quite enough for constructing and implementing
motion of snakelike robot with only two DOF. Numerical researches also show
that the width of “corridor1‘(deviations of a system from pre-set trajectories) is
sufficiently small, and more, can be controlled and may be reduced due to an
optimum choice of the parameters which setting curvature of turns. It has been
also numerically investigated behavior of system under conditions of
disturbances and it was shown the stability of the robotrain system in sufficient
range of noise (disturbances) values. These properties should be considered by a
control system realizing movement of real “robotrain”.

REFERENCES

1. F. Lamiae, J.P. Almond. “A practical approach to feedback control for

a mobile robot with trailer” in Proc.1998 IEEE International Conference on
Robotics & Automation, pp. 3291-3296, 1998.
2. Y. Nakamura, Hezekiah, Yutan, W.Chung. “Design of Steering
Mechanism and Control of Nonholonomic Trailer Systems” in Proc. 2000 IEEE
International Conference on Robotics & Automation, pp. 247-254,2000.
3. K.-U. Scholl, V. Kepplin, K. Berns, R. Dillmann “Controlling a
Multijoint Robot for Autonomous Sewer Inspection’’ in Proc. 2000 IEEE
International Conference on Robotics & Automation, pp. 1701-1706,2000.
4. M.Vendittelli, G.Oriolo. “Stabilization of the general two-trailer
system” in Proc. 2000 IEEE International Conference on Robotics &
Automation, pp. 1817-1823,2000.
5. V.E.Pavlovsky, N.V.Petrovskaya. Investigations of dynamics of
“robotrain” chain motion. Equations of motion, specific solutions. /I Preprint of
KIAM, 2005, 117. In Russian.
6. V.E.Pavlovsky, V.V.Evgrafov, N.V.Petrovskaya. Investigations of
dynamics of ”robotrain“ chain motion. Motion under control. // Preprint of
KIAM, 2005, 120. In Russian.
7. V.E.Pavlovsky, N.V.Petrovskaya. Investigations of dynamics of
‘‘robotrain’’chain motion. Motion planning techniques. // Preprht of KIM,
2005, 121. In Russian.
8. Fukushima, E.F. and S . Hirose; Optimal Attitude Control for Articulated
Body Mobile Robots. I/ Proc.of the International Symposium on Adaptive
Motion of Animals and Machines, CD-ROM proceedings, Montreal, Canada,
August (2000).
 S E R V I C ~ GSOLAR POWER PLANTS
WITH W A L L W A L ~ R

RIDHA AZAIZ
R WTH Aachen, Wallwalke~. Net
Kastanienweg 6, 52074 Aachen, Germany

More and more solar power plants are being installed worldwide in concern over climate
change. The efficiency of renewable energy sources is crucial as the power output is
being sold. Deposits, dust and damages lower the output of the solar plant. This is
usually not recognized by the operator of the plant, as they are installed at a high altitude
and inclination. For economic reasons and safety at work, inspection and cleaning of the
solar plant is oRen neglected. This is where Wallwalker applys. Wallwalker is a light
weight robot that is able to walk on the sensitive solar modules in order to inspect and
clean them. The cleaning process is controlled by an imaging software. In addition
damages can be detected and are reported to the operator. The efficiency of the solar
power plant raises and mortises faster.

1. Introdu~tion

1.1. Concept of Solar Power Plants

Solar power plants consist of several modules that are switched together. In
order to gain the highest output, they are usually ~ n s ~ with a ~ an
~ ~inclination
d
tonvards the sun.
Each module consists of many solar cells. These singie cells are switched
serial into a string, and the strings are again switched together. The solar module
is protected by glass or lamination, embedded in an aluminium frame.

Figure 1. Solar power plant consisting of several modules on different sites

219
220

1.2. How deposits effect the plan1

If one solar cell of the module was shadowed, e.g. by deposits such as dust, it
would produce less current than the cells that are switched before and after. The
cell with the lower output would actually consume current, heat up and damage.
Therefore the manufacture of the module install diodes, in order to switch
the eRected cell of. But as each diode needs current itself, manufactures
compromise on integrating only one diode into each string. That means that in
the case of deposits on one cell, the whole string gets disconnected. The output
of the effected module declines in high proportion.

disconnected string
solar cell with deposits

Figure 2. Solar module consisting of several cells, one with deposits.

2. A u t o ~ a t e dcleaning and inspection with i all walker

2.1. ~ e ~ ~ i r e ~oferobotic
n t s se~vi~in~
As solar modules are usually installed at high altitude and inclination, they are
hard and dangerous to reach by manpower. Additional equipment and
i n f r a s t ~ c ~irs eneeded to clean them manually. The load of the human body is
often too heavy for the modules. Limiting the stress during servicing by walking
on the aluminium frames only remains difficult.
A criteria in automated cleaning of solar modules is light weight of the
device. It is also crucial that the robot is not damaging the surface while cleaning
and moving, since that would have effects on the output of the plant, too. The
plant consists of several modules that have spaces and obstacles such as
mount in^ materials in between. The robot needs to be able to recognize and to
overcome such obstacles.
Picture 1 indicates that the robot should be able to move on surfaces at an
inclination of up to 90 degree and a surface t e ~ p e r a ~ of r e up to 100 degree
Celsius. The robot should be able to reach every point on an rectangle shaped
surface. It is also an advantage if the robot can be moved from one module-site
to an other easily without additional installation.
 22 1

2.2. C ~ a r a c ~ofe ~~ a~~~~ ~w ~

a ~s k e r
The specified requirements led to the development of Wallwalker. This is a
robot that moves back and forth and takes turns in order to cover a whole
rectangular surface. Wallwalker is a wall climbing robot that attaches on the
surface with vacuum bulbs at an inclination of up to 90 degree.
The light weight (6kg) ensures that the surface is not damaged by the robot
and the material of the vacuum bulbs is resistent towards the temperatures on the
surface. The framework of the moving unit makes it possible to step over
obstacles, such as mounting materials, with a width of up to 8 cm and a height of
3 cm. A supply secures the robot and carries both electricity and cleaning agent.
The camera compares the surface before and after the cleaning process.

Figure 3. Wallwalker on solar modules.

Figure 4. Front view of WallwalkeI

222

2.3. ~ o c o m o ~ i o ~
Wallwalker moves back and forth on a so called “sliding-frame”. This frame
consists of two parts. One is the inner part that has its own vacuum bulbs and
moves up and down. The other is the outer part that moves the inner part back
and forth. First, the robot moves its inner part towards the surface and attaches
the vacuum bulbs onto it. Once that happened the outer part is lifted and moves.
Now the outer part can be lowered and attaches onto the surface. When the
vacuum bulbs are evacuating air, the inner part is detaches and lifts. Next, that
part moves and the steps are repeated.

Figure 5. Simplyfied scheme of Wallwalker walking.

 223

2.4. Covering s t r u t e ~
Solar power plants usually have a relatively simple shaped surface. A rectangle
can be covered most easily by the robot as it moves like a snake. This means
Wallwalker moves forth until it reaches an edge which it can not step over. This
could be the edge of the surface. There it turns 90 degrees to the right, steps its
own with forth and then turns again 90 degrees. Now it reached its next track and
steps forth until the next edge it can not step over (that would be the end of the
surface). Wallwalker will turn left there and repeats the steps it did before.

Figure 6. Wallwalker covering a large, rectangle surface

2.5, Services
The efficiency loss through deposits is strongly dependend on the geographic
site of the plant. In regions like Spain, California and Africa where there is less
rain deposits stay on the glass longer. Rain does not clean everything. Dust and
polls bum into the glass and become hard to clean off.
Nanostructures are not imposed on solar modules for several reasons. One
of them is that additional surfaces on the glass will higher the reflexion of the
sunbeam, lowering the efficiency of the solar modules again.
Besides this other environmental circumstances lead to corrosion,
delamination and other damages. These errors are hard to identify and to keep
track of in the large solar power plants. Therefore Wallwalker has a cleaning
module attached and a camera. The cleaning unit consists of a rotating
microfibre towel that is supplied with the cleaning agent. The camera and the
imaging software compare the result of the cleanig process before and after.
224

Figure 7. Imaging software with clean reference cell

Figure 8. Solar module with deposits

 225

2.6. Summury and outlook

Wallwalker is a mobile robot that ensures the efficiency of solar power plants.
With the robot surfaces of high altitude and inclination can be maintained where
it has been too dangerous and expensive before. Wallwalkers' simplicity and
light weight offer a unique and broad range of applications for measurement,
construction and maintenance purposes.

References
1. H. Haberlein, D. Graf Gradual Reduction of PV Yield due to
Pollution, 2"d World Conference on Photovoltaic Solar Energy
Conversion, 1998
2. C.Ranken, H. Haberlein: Langzeitverhalten von netzgekoppelten
Photovoltaikanlagen, Ingenieurschule Burgdorf, 1 1/99
 STABILITY AND GAIT OPTIMIZATION OF A HYBRID
LEGGED-WHEELED ROVER

BYRON E. JOHNS
School of Mechanical Engineering
Human-Automation Systems Lab
Georgia Institute of Technology
Atlanta, GA 30332, USA

AYANNA M. HOWARD
School ofElectrica1 and Computer Engineering
Human-Automation Sysiems Lab
Georgia Institute of Technology
Atlanta, GA 30332, USA

In this paper, we discuss the stability and gait optimization of a new kind of
hybrid legged-wheeled rover. In order to control this complex robot, kinematic
analysis is needed for both its wheeled and legged mobility system. With these
kinematic models, we analyze the proper location of the legs during both the
standing and walking phase and develop an optimization scheme to provide a
stable locomotion sequence for the robot when in its legged configuration. We
also compare and contrast two walking gaits associated with hexapod robots,
and converge on a final gait based on this analysis for the legged mobility
system of this new hybrid rover. Details of the optimization approach as well
as results using the physical hardware are presented in the paper.

1. Introduction
Exploration on unknown and uncharted planetary surfaces typically involves
operating in an unstructured and poorly modeled environment. To allow robust
navigation capability in these natural environments, several designs are plausible
for such a mobile robotic platform [ 11. In order to guarantee success of robotic
missions for the future, we focus on modularizing both hardware and software
components to create a reconfigurable robotic explorer. This allows the
deployment of rovers on a planet’s surface that can ensure robust operation to
deal with changes in the face of system reconfiguration, hardware failure,
changes in task specifications, or alterations in the terrain environment. We have
developed a new reconfigurable legged-wheeled rover known as Byrobot, which
capitalizes on the positives of both wheeled and legged locomotion. Byrobot has
a six-leg hexapod configuration with a 3-revolute joint mechanism on each leg as

226
 227

well as 4 DC motors controlling the four wheels on the robot. In this paper, we
discuss ow methodology for stabilizing the robot in its legged configuration as
we develop an effective walking gait. Figure 1 shows CAD models and actual
photos of the Byrobot platform.

Figure 1:CAD model and actual hardware photos ofByrobot

2. Stability Analysis of a Legged Mobility System

For a legged robot in remote terrain environments such as Mars, stability is a
more important issue than speed. If the robot can’t remain stable and stand
upright on its own, then the objectives of the mission can not be achieved. A
wheeled mobility system has inherent stability characteristics since the wheels
are usually always in contact with the ground. However, there are areas such as
cliffs, rocky, and soft terrain, which a wheeled robot may not be able to stabilize
on and traverse. For a hexapod robot, at least 3 legs have to be down at all times
for the robot to be in a stable configuration. At any given moment, the robot’s
center of mass (COM) should be within the “Polygon of Support”, meaning that
the center of mass is as far away from the edges of the polygon [2] that covers
the area defined by the feet on the legs that are in contact with the ground. .

3. Forward and Inverse Kinematics

In order to position the robot feet for attaining proper stable polygon of support,
forward and/or inverse kinematics is used [3]. Here we approximate where the
center of mass (COM) is located for simplicity.

3.1. Forward Ki~ematics- Geometrical Method

The simplest method of forward kinematics calculates the foot position based on
a geometric representation of the legs in the x, y, z coordinate frame, as we see
in Figure 3. This method allows us to set the joint angles of the robot legs and
determine the final position of the foot [4].
228

I I
I pelvic joint I

I
I
I
I
I I
I
; ~ 3 1 end effector position
I
I Y (foot) I

Foot Position
x3= Llcos(BI)+ L2cos(B2)sin(B1)+ L3c~~(e2+B3)cos(B1)
Y3 = Llsin(BI)+ L2~~~(B2)cos(B1) + L3cos(B2+O3)sin(Bl)
Z3 = L2sin(B2)- L3sin(Bz+B3)- d2

Where L],,Lz,and L3 are the leg segment lengths, el,€I2, and B3 are the pelvic, hip,
and knee joints of the leg mechanism, respectively, and d2 is the offset length in
the z-direction from the leg origin down to the L1 leg segment, all shown in
Figure 3.

3.2. Inverse Kinematics

The inverse kinematics can be used determine the joint angles (our 0’s)
associated with the legs, given the position of the foot. This is important for
the standing and the walking of Byrobot. A single set of three joint angles
produces a single foot position; however, there is more than one set of joint
angles that will result in a given foot position (as you can see by planting
your foot on the floor and moving your hip and knee). Therefore,
incongruous solutions, such as a knee bending backward, must be
eliminated. The inverse kinematics are derived from [ 5 ] , and implemented
in a gait simulation for a similar 3-DOF hexapod in [6].
 229

Where we use the same variables as mentioned above.

4. A Special Case -Robot on an Incline/Decline

In the case when the robot is on an incline or decline, the projected center of
mass (COM) is into the surface plane, as shown in Figure 4.

Figure 4: ~ y r o bprojected
o~ down on an incl~ne/decl~ne

The COM is projected into the plane at cos(0), where 0 is the angle of the
incline. In this situation, the walking gait needs to be modified so that the COM
doesn't cause the robot to tip over, This modification can be done by adjusting
the joint angles in the robot legs to correspond to the incline angle.

5. Walking Gait Optimization

We tested two walking gaits that are commonly used for hexapod robots [XI.
Each gait depends on the load on the robot, how fast you want the robot to
move, and how stable the robot will be while moving.

5.1. The Tr@d Gait

The Tripod Gait is the best-known hexapod gait. A tripod consists of the front-
back legs on one side and the middle leg on the opposite side. For each tripod,
230

the legs are lifted, moved forward, and lowered in unison. Since 3 legs are on the
ground at all times, this gait is both “statically” and “dynamically” stable. This
gait uses two leg movements, or two “tripods” to complete one cycle, making it
the fastest gait for hexapod robots.

5.2. The Wave Gait

In the Wave Gait, all legs on one side are moved forward in succession, one
after the other. This is then repeated on the other side. Since only 1 leg is ever
lifted at a time, with the other 5 being down, the robot is always in a highly-
stable posture, making this the most stable gait and the preferred gait for robots
with large loads. However, since this gait takes six leg movements to complete
one cycle, it is also the slowest gait.

5.3. Movement Scheme

The movement scheme is easily visualized by examining Figure 5. The square
wave graph next to each of the six robot legs correspond to time points. In this
figure, Hi means the leg is lifted off the ground, while Lo means the legs is down
on the ground.

TriDod Gait - lift 3 lees at a time (3 lees down)

IHi
n -LllnIL
Hi
In

Wave Gait - lift 1 lee at a time (5 lees down)

Hi
In
n Hi n
Hi
In
n Hi

Figure 5: Different Gait movements of each leg with respect to time.

 231

6 . Results
Our goal is to stabilize our new legged-wheeled rover and test two different
walking gaits for its hexapod configuration. We have shown the kinematic
models for the leg mechanisms, explained how the polygon of support is needed
to stabilize the robot in its legged configuration, and we have tested out the
Tripod Gait and the Wave Gait. Due to the large load that is on the Byrobot
(approximately 2kg), the initial tripod gait was ineffective. Byrobot simply
would collapse under its own weight as soon as the three legs would lift off of
the ground, despite the leg position. The Wave Gait, though a bit slow, remains
stable throughout its entire walk, and therefore became our preferred walking
gait. The correct servo joint angles were found using the inverse kinematics, for
each step in the gait. For the Wave Gait, the movement of the six legs is in the
sequence of: leg 1, leg 3, leg 2, then leg 4, leg 6, and leg 5 (robot legs labeled in
Figure 5), so that there was always a stable polygon. At the end of each cycle
during the gait when all six legs have moved a forward, a "forward scoot" is
done to move the robot body forward. Small steps were used because they are
more efficient, and we wanted to avoid interference of parts on the robot that
could occur if we used larger steps. In Figure 6, we see photographs of the
successfkl Wave Gait sequence of Byrobot on a flat surface.

Figure 6: Photographs of the Vave Gait sequence of Byrobof

In addition to the flat surface, we also tested Byrobot on an incline plane and in
our Mars Sand Pit. We compared the results of both the legs and wheels in these
environments. For our incline plane, we used 22" and 35" ramps. A 35" angle is
the maximum angle the robot can traverse on its wheels before slipping back
down. U n f o ~ a t e l yon the incline plane, the robot would not remain stable
using the on its legged mobility system, making the wheels the preferred choice
for mobility. In the sandy terrain, the wheels would lose traction, making the
legs the preferred mobility system. For operation in this terrain, Mars Shoes
were used on Byrobot. These shoes give more surface area to the feet so that
they don't dig down into the sand. In Figure 7, we see the Mars Shoes in place
on Byrobot and Byrobot standing in the shoes in the Mars Sand Pit.
232

Figure 7: Mars Shoes on thefeet of Byrobot

The average speed of the robot using its legs in the sandy terrain was about half
of the speed of it using its legs on a flat surface. We contributed this to the lack
of friction on the robot's Mars Shoes which cause its feet to slightly slip
backward with each step during the walking gait. As expected, the legs move the
robot much slower overall than the wheels do. Our average speed results are
shown in Table 1.

Up an Incline Plane
Wheels * 22" - Full Speed 37.56
0 35" - Full Speed
Legs e Full Speed nla

Sandy Terrain
Wheels I * Full Speed n/a
Legs I e Full Speed 0.54

7. Conclusions
In this paper, we have explained how to stabilize a hexapod rover using forward
and inverse kinematics to find its polygon of support. And we have shown how
 233

to find an optimal stable walking gait based on these kinematics. We tested two
of the common walking gait patterns associated with hexapod robots and
explained why our final Wave Gait was chosen. We will be able to improve the
mobility of the robot on an incline plane in the future by adjusting the joint
angles. As shown in Figure 4, the center of mass of the robot is projected down
into the plane, and the legs will need to be in a different configuration depending
of the angle of the incline, in order to keep the robot stable in this situation.
Having a robot that can both walk and roll would be beneficial to planetary
exploration since it can transition from the use of its legs or wheels based on the
terrain conditions. By adding various sensors, encoders, and even a vision
system, we can facilitate future autonomous capabilities of Byrobot.

References
[ 11. Howard, A., Tunstel, E. “Intelligence for Space Robotics”. TSI Press,
2006
[2]. Quinn, R., Nelson, G., Bachmann, R., Kingsley, D., Offi, J., Ritzmann,
Roy. “Insect Designs for Improved Robot Mobility”. 4th International
Conference on Climbing and Walking Robots, 200 1.
[3]. Craig J.J. Introduction to Robotics, Mechanics and Control, 3rd ed.
Addison Wesley, 1986, 1989,2005.
[4]. Sciavicco, L. and Siciliano, B. Modeling and Control of Robot
Manipulators. McGraw-Hill Co., Inc. 1996,2000.
[5]. J. Lento, Z. Huson , J. A. Haass, M. Reilley, and J. Shrestha.
Development of Leg Control Mechanisms for a Radially Symmetric
Octopedal Robot. The National Conference on Undergraduate
Research, 2005.
[6]. G. Figliolini, V. Ripa, (2004) Kinematic Model and Absolute Gait
Simulation of a Six-Legged Walking Robot. CLAWAR 2004 889-896,
2004.
[7]. Ferrell, C. “Robust and Adaptive Locomotion of an Autonomous
Hexapod”. Perception to Action Conference, Lausanne, Switzerland,
66-77, 1994.
[8]. NASA JPL Robotic Systems; http://www-robotics.ir,l.nasa.gov
 TERRAIN-ADAPTIVE LOCOMOTION OF A WHEEL-LEGGED
SERVICE ROBOT USING ACTUATOR-BASED FORCE
MEASUREMENTS

PETRI VIREKOSKI
Automation technology laboratory, Helsinki University of Technology, Otaniementie 17,
0201 5 TKK, Finland

ILKKA LEPPANEN
Automation technology laboratory?Helsinki University of Technology, Otaniementie 17,
02015 TKK. Finland

1. Introduction
This paper describes methods for adapting to terrain with a service robot that has
four legs in a particular configuration that includes wheels at the ends of the legs.
The basic motivation behind this type of configuration is to use the wheels to
give more speed on terrain that allows it and use the legs more when the terrain
becomes more difficult or impossible to negotiate using wheeIs only.
Terrain adaptation here means active utilization of the legs to gain the best
possible support and propulsion in varying and rough terrain, and therefore also
means trying to maintain equal ground contact with all four wheels. It acts in part
the same way as the spnng/dampener wheel suspension in cars and gives the
following benefits: equalization of strain forces, better propulsion in soft soil,
locomotion beyond the ability of wheels and also less strain to the terrain.
The terrain-adaptive locomotion is based on measuring forces and
controlling positions of the legs’ ankles (and thus the wheels). Furthermore the
force measurements are not based on external sensors but instead the forces are
calculated directly using the currents of the electric actuators (motors) and the
kinematic equations of the robot.

2. The wheeled leg configuration

The robot used is the WorkPartner service robot that has been developed in the
Automation Technology Laboratory of the Helsinki University of Technology. It
has four legs in a particular configuration, and wheels at the ends of the legs. The

234
 235

con~gurationhas been described in previous CLAWAR papers [I] and [ 2 ] .

Figure 1 shows W o r ~ a ~ ine roughr terrain. Figure 2 shows the degrees of
freedom of the leg. The main body consists of two parts, front and rear, joined by
a scissors-type middle joint for steering. The front legs are connected to the front
body and the rear legs to the rear body.

Figure 1 WorkPartner negotiating some obstacles in winter

Figure 2 The degrees of freedom of the leg of WorWartner

It should be noted here that while the legs have the “hip” joint (longi~dinal
axis, joins the leg to the body) for sideways ankle movement, this joint is not
236

used for the methods described here. So only the 2nd and 3rd (“thigh” and
“knee”) joints are used to move the ankle in the vertical lengthwise plane.

3. Modes of Locomotion
First let us take a look at the different main modes that are determined by how the
legs and wheels are used for locomotion. The three possibilities are: wheels only,
hybrid mode (use of both the leg joints and wheels) and walking (use only leg
joints, wheels locked). The term “rolking” (combining the words rolling and
walking) will be used from here on to refer to the hybrid mode.

3.1. Wheels-only Drive

This basic mode uses the wheels in velocity mode to drive forward and back, and
the middlejoint is then responsible for steering. This mode is similar to tractors
that use a hydraulic middlejoint. The leg joints can be either completely fixed, or
they can be used for adapting to the terrain. The “wheel” mode is best for mostly
flat terrain, allowing greater speeds because the legs do not have to participate in
moving the vehicle forward and back.

3.2. Rolking (Rolling-walking)

The “rolk” mode combines wheel rotation and horizontal leg motion to handle
difficult terrain in such a way that the wheels can retain ground contact all the
time. It is similar to walking except the legs are not lifted off the ground. This
helps with a) stability problems that are related to walking where ground contact
is lost periodically and b) wheel slipping problems related to normal wheel drive.
While rolking the legs can have one of two states: transfer or support. In
support state the leg participates in moving (pushing) the vehicle body forward
while the wheel is kept stationary. In transfer state the leg is moving forward or
backward while the wheel is rolling along the ground.
When only one leg is in transfer state at a time, there are always three legs in
support state pushing the vehicle forward. The transfer leg’s wheel is kept in
velocity mode where the velocity is exactly the same as the ankle’s motion in
relation to ground. This helps particularly on slippery or delicate surfaces,
’ because there are no wheels that could slip (unless the whole vehicle slides).

Rolking can also be done with two wheels in transfer state simultaneously.
This gives more speed because there are only two separate phases. Particularly
the diagonal two-leg rolking has been found effective with WorkPartner in soft
soil.
 237

The rolk mode is also suitable for climbing over obstacles because only one
leg is moved at a time, allowing simpler control over how the legs behave when
climbing on the obstacle.
Steering is not currently possible while rolking. This is because changes to
the middle joint angle would require changes to the legs’ positions in order to
keep the wheels from sliding sideways, and this would require rather complex
adaptation for the rolking algorithm or might even be impossible to implement
without sliding of the wheels. So the middle joint is kept locked while rolking.

3.3. Walking
Four-legged walking requires complex control of dynamics and stability. The
terrain-adapting methods described later have not been developed for walking
and are not applicable to it, although later in this paper we see that the “rolking”
mode could lead to walking when using a particular scheme for adapting to
terrain and obstacles.

4. Adapting to Terrain
When a vehicle has four wheels at fixed positions, and is moving along a terrain
that is no longer flat, one wheel loses contact with ground causing essentially two
problems: 1) that wheel can no longer participate in propelling the vehicle and 2)
the weight of the vehicle is divided between the remaining wheels causing more
strain both on the wheel suspension and the terrain. When this happens,
depending on the location of the center of mass, it ultimately leads to only two
diagonal wheels bearing most of the weight and leaving the other two almost
without grip.
In a car this is handled by suspending the wheels so that they can move a
distance vertically and putting springs and dampeners between the body and the
wheels to cope with the weight and mass. In the case of our robot, the wheels are
“suspended” by legs that have greater vertical reach but can also move the wheel
horizontally. This gives more flexibility to handle uneven terrain, but also means
that all terrain adaptation is dependent on active measuring and actuation.
The control of terrain adaptation is based on measuring forces and changing
position. That is, the position of the ankle of each leg is varied vertically
depending on what kind of strain is present at the ankle. For a simplified
example, if a leg feels “too light”, it is pushed “a bit” more down to try to regain
proper load; What exactly is “too light” and how much is “a bit”, are parameters
of the tuning of the control.
238

Force control in walking robots has been studied in e.g. [3] and [4] where
the goal has been to divide the support forces between the legs according to a
certain principle such as to equalize the vertical forces or minimize internal
forces. These have required rather complex algorithms. In a four-legged system
using only the “wheel” and “rolk” locomotion modes the support force
equalization can be done using simpler methods described here.
Figure 3 shows an extreme situation where one leg is driven on a large
“bump” (in the “wheel” mode), and the terrain adaptation can be seen clearly.

Figure 3 An extreme case of tenain adaptation

5. Controlling Terrain Adaptation in Workpartner

Both the “wheel” and “rolk” modes use a same scheme for controlling the legs.
The horizontal and vertical coordinates of the ankle positions are controlled
separately:
* The ho~izontalposition is determined by the locomotion mode. When
rolking the gait sequencer moves the ankles horizontally by a pre-
determined sequence (each leg in turn), and in wheel mode the
horizontal position is fixed.
* The vertical position is controlled by the terrain adaptation algorithm.
This means that the basic terrain adaptation is independent of the main
locomotion mode.
Terrain adaptation can have two modes, “level” and “terrain”. In “level”
mode the controller attempts to keep the body perpendicular to gravitation (using
 239

two inclination sensors). In “terrain” mode the leg positions are adjusted so that
the body follows the local terrain profile. The control loop can be divided into
three steps: 1) initial positioning according to adaptation mode (this includes the
leveling control in “level” mode), 2) equalization of the vertical forces and 3)
averaging and limiting.
The forces that are used as inputs for the terrain adaptation are calculated
from the electric motor currents in the “thigh” and “knee” joints. The currents are
converted to motor torques and those are converted to joint torques. Using the
kinematic equations of the legs the joint torques are then converted to vertical
and horizontal force vectors in the robot’s coordinate system. These are the actual
strain forces acting on the ankle.

Step 1: Initial Positioning, Including Leveling Control

Here the initial vertical positions are set according to the adaptation mode. In
“level” mode this step adjusts these positions so that the body keeps level
according to gravitation and a desired height from the operator. In “terrain” mode
the positions are set according to desired pitch, roll and height of the body from
the operator.

Step 2: Force Equalization

Now the calculated vertical strain forces of the ankles are used to determine an
adjustment to the two diagonals. This means that the legs are moved together in
the two diagonal pairs (legs 1&4 and 2&3, see Figure 4) to adjust the division of
forces. The division is determined by comparing the two sums of the vertical
forces of each diagonal pair. In “wheel” mode the sums are kept equal. When
rolking the adjustment is weighted so that the transfer leg’s diagonal is lighter i.e.
the sum of the two diagonal legs’ vertical forces is less than the sum of the other
two. This is done in order to reduce the rolling resistance of the transfer leg’s
wheel.

/I
II
Figure 4 The numbering of the legs (wheels)
240

Step 3: Averaging and Limiting

When in “terrain” mode the positions are averaged between the left/right and
fronthear pairs. First the mean position of the left legs is compared to that of the
right legs, and both groups are moved to opposite directions by half of the
difference. Then the same is done for fronthear legs. This averaging is not done
in “level” mode because it would defeat the leveling. Finally the positions are
limited to tested practical maximums, and set for sending to the legs.

An important point to note is that in steps 2 and 3 the legs are always moved in
opposing groups of two, and to opposite directions by the same amount, so that
the average height is maintained.
In step 2 the diagonal lightening during rolking can be increased even to a
degree where the rolking (transfer leg) diagonal is completely without load. This
has the effect of changing the rolking to walking, and has been confirmed in tests.
However with the static implementation of the rolking gait this walking behavior
is not completely predictable because depending on the projection of the center of
mass either the rolking leg or its diagonal opposite can lift up. This has not been
studied further.
Figure 5 shows a test that clearly indicates the effects of force equalization
(see caption).

Figure 5 A test where leg 2 (front right) was driven over a 1 m long and 20 cm high smooth bump
without (left) and with (right) terrain adaptation, in wheel mode. As expected, the support force
variation is greatly reduced with the force equalization in terrain adaptation (“terrain” mode was used,
see text for description). The ripple is caused by the relatively low frequency of the force equalization
controller combined with mechanical flexibilities and play in the joints.
 241

6. Conclusions and Future

The terrain adaptation method and algorithm described here has been widely
tested in varying terrain including dirt, sand, snow and obstacles like logs on the
ground. The following benefits have been confirmed:
requires no additional sensors
works as well in both the “wheel” and “rolk” main locomotion modes
divides the weight equally in “wheel” mode which leads to
o better propulsion
o less rolling resistance in soft soil
o less strain on both robot and terrain
0 works together with the basic motion control of the legs
0 gives mobility when just wheels would not suffice (e.g. deep snow)

A couple of future enhancements have been proposed and partly tested for the
terrain adaptation. First, the robot could switch the locomotion mode
automatically based on measuring the vehicle-terrain interaction. This has
actually already been developed quite far and is working well. Second, it would
be possible to control the projection of the center of mass to some extent by
changing the global lengthwise and sideways position of the wheels relative to
the body (this would also utilize the “hip” joints). This would give even more
control to the equalization of the strain forces, particularly in inclined terrain.
A study has been started to examine the possibility of using similar active
wheel suspension and terrain adaptation to improve mobility and payload
handling in such machines as tractors and loaders.

References
1 . Leppanen I, Salmi S, Halme A (1998) WorkPartner - HUT-Automations
new hybrid walking machine. 1st International Conference on Climbing and
Walk-ing Robots, Brussels
2. Halme A, Leppanen I, Salmi S (1999) Development of WorkPartner-robot -
design of actuating and motion control system. 2nd International Conference
on Climbing and Walking Robots, Portsmouth
3. Gorinevsky D M, Shneider A Y (1990) Force control in locomotion of
legged vehicles over rigid and soft surfaces. The Int. J. of Robotics Research
VOI 9. NO. 2 ,p. 4 - 23
4. Lehtinen H (1994) Force based motion control of a walking machine. Ph. D.
dissertation, Helsinki University of Technology
 Control of Quadruped Walking Robot
Based on Biologically Inspired Approach

Tae Hun Kang

Phohang Institute of Intelligent Robotics, Korea
E-mail: bigxihn@postech.ac.kr

Ig Mo Koo, Young Kuk Song, GiaLoc Vo, Tran Duc Trong, Chang Min Lee
and Hyouk Ryeol Choi'
School of Mechanical Engineering, Sungkyunlnuan University,
Chunchun-dong 300, Jangan-gu, Suwon, Kyunggi-do, Korea,
E-mail: kooigmo ; yksong99 ; vogialoc ;jamestran ; cmbof ; * hrchoi{ @me.skku.ac.kr}

In this research, a biological study is performed on the control of a quadruped

walking robot. With a biomimetic observation of the gravity load receptor and
stimulus-reaction mechanism of quadrupeds' locomotion and with the study
of the stances on walking and energy efficiency, a new control method for
a quadruped walking robot is derived. In particular, introduction of a new
rhythmic pattern generator relieves the heavy computational burden because
it does not need large computation on kinematics. The proposed controller,
though it is simple, provides a useful framework for controlling a quadruped
walking robot. In this paper, evaluation of the proposed method is done via
a dynamic simulation. Then, it is validated by implementing in a quadruped
walking robot, called AiDIN(Artificia1 Digitigrade for Natural Environment).

Keywords: Biomimetic; Gravity Load Controller ; Quadruped Walking Robot

1. Introduction
In spite of rapid development in robotic technologies, living creatures are
superior to robots existing currently in various aspects. Thus, it is necessary
to understand the principles underlying the motions and behaviors of bio-
logical subjects for innovative control of robots. Mimicking living creatures
currently becomes one of the worldwide trends for robotic innovations. It
is considered as one of the most adequate way of developing a robot since
biological systems provide a number of useful ideas concerning control of
robots. Recently, robotic researchers as well as biologists propose innova-
tive ideas for the walking robot system. Among several ideas, mimicking

242
 243

Motor driver
/ SBG(RTLinux)

Force sensor

Fig. 1. AiDIN( Artificial Digitigrade for Natural Environment)

the rhythmic motion of animals is one of the promising ways to control the
walking system. By studying on this, the locomotion of the walking robot
can be close to that of the real animal. Recently, neurobiologists point out
clear evidence of a neural part in the brain of animals, referred as Central
Pattern Generator (or CPG), which can produce the rhythmic movement in
its locomotion. Since CPG can generate rhythmic outputs even if the sen-
sory feedback is eliminated, it became an interesting target in approaching
to biomimetic control. On the way to study about this CPG, many neural
models such as McCulloch and Pitts neuron, Leaky integrator neuron, and
Matsuoka neuron have been proposed by investigating neurons in the real
animal’s brains or bodies [1,2]. As efforts of developing relevant methods to
control the robots, algorithms inspired from these rhythmic patterns gen-
erators have been studied [3]. Though these neural models can illustrate
the control method of animal movements, the results of applications in
robotic systems are far different from real biological systems because ani-
mal’s movement may not be simply mimicked by several neurons as robotic
systems introduced until now [3].
On developing a quadruped walking robot, various aspects should
be taken into account. In this research, mainly, the basic principles of
quadrupedal locomotion controller are analyzed in terms of a biomimetic
point of view. The quadrupedal locomotion controller is studied via
biomimetic observation of the gravity load receptor and stimulus-reaction
mechanism of animal are investigated using the stances on walking and
posture control. The controller design factors is feedback information form
244

several receptors and its stimulus-reaction based on biological sensory sys-

tem analysis. Finally, the proposed idea is applied in a quadruped walking
robot, called AiDIN( Artificial Digitigrade for Natural Environment) , and
the effectiveness of the proposed idea is experimentally validated.

2. Gravity Load Controller

Usually, animals move by the actions of muscles. When muscles generate
arbitrary movements, the control action is executed via two loops. In the
first loop, a command from the brain is transmitted to the muscle through
a central nervous system(CNS). In the other loop, the muscle reacts to ex-
ternal stimulus which are not predictable or recognizable. The stimulus are
detected with receptors such as proprioceptor and exteroceptor [4].Even
though its structure is very complex, movements can be easily simplified by
the repetitive action such as flexion and extension, because these actions
are directly controlled by sensory neurons and motor neurons [5]. In this
section, a locomotion controller, called Gravity Load Controller(abbreviated
as GLC afterwards) is introduced. This controller is based on a gravity load
receptor and stimulus-reaction mechanism of the animal. Even though the
configuration of GLC depends upon the system in hand, the basic configu-
ration can be shared because it consists of the feedback information from
several receptors and its stimulus-reaction.
As shown in Fig. 2, GLC for the quadruped walking robot consists of
four parts including an oscillator, stimulator, threshold detectors, and PD
controller. In addition, the proposed controller has two loops( exceeding and
nonexceeding) which are separately excited according to the threshold de-
tector. If the information from sensory feedback(it is called load receptor in
biology) such as force sensor, touch sensor, and gyroscope exceeds the fixed
threshold value, the stimulator begins to excite the extensors in the dotted
rectangle. Then, intennittent stimulus are generated, which propel a joint
to react immediately to external feedback. In the nonexceeding case, the ex-
tensors in the dotted rectangle are disabled, and the stimulator stimulates
the flexor and extensor which are connected with 6c called interneuron. 6c
is within the central nervous system that acts as a link between sensory
neurons and motor neurons. Two interneurons, 6c and 61 in the controller
are expressed with simple equations as follows.

where
 245

Kg
Oscillator
Proprioceptor
Gyroscope I, I Continuousstimulus
information

Intermittent stimulus
Force
information
Exteroceptor Threshold 8 Stimulator

Fig. 2. The joint controller consists of four parts including an oscillator, stimulatory
part, threshold detector, and PD controller

.={ 1 otherwisewith
-1
contact ground
in supporting mode

€ ={ -1 otherwisewith ground
1
contact
in swing mode

Forelimb I Hind limb

ST I SW I ST 1 SW

Also, a new oscillator is proposed, which is simple and easy to implement

in the locomotion controller. Although many researchers have investigated
the ratio between SWand ST according to the walking speed, it is very
difficult to apply this idea to the robot system. A ratio between S W and
ST needs to be normalized to decide a general walking pattern for a cycle.
From several references [6,7], digitigrade’s walking patterns can be mea-
sured, which are normalized as shown in Table l. From these observations,
it is noted that ratio between SW and ST is decided by walking speed as
246

well as walking types, which is applied in the design of the oscillator.

The proposed oscillator iCij(joint j of leg i) is constructed with Bezier
curve with n control points. The oscillator K i j can be represented as the
function of time t E [0,1] such as

where PI, is a control point.

From Eq. (3) the oscillation pattern for the quadruped walking robot
can be designed. The desired joint angle eij of leg i and joint j is given by
the stimulus of extensors, interneurons, and oscillator as follows.

An oscillator Kij in the proposed controller generates a repetitive pat-

tern for locomotion. i C i j always generates locomotion pattern regardless of
any input value as illustrated in Fig. 2. It is quite different from the other
pattern generator introduced until now.
As illustrated in Fig. 3, the overall controller for the quadruped walking
robot is comprised of four joint controller and the motion is synchronized
with the switch for determining the walking patterns.

I/ TP

Fig. 3. Illustration of the walking controller

 247

3. Quadruped Walking Robot: AIDIN

A quadruped walking robot, called AiDIN ( A ~ ~ ~ c~ i~agli t i g rfor
a ~~e a t u r a l
~ n v ~ r o n ~illustrated
e n ~ ) in Fig. 1. AiDIN is the robot containing all the
components except the battery AC 22OV power is supplied via a tether
cable.

(a) Forelimb (b) Hindlimb

Fig. 4. Schematic view of limbs

As shown in Fig. 1, AiDIN has four legs with three-DOF active joints
for each leg. The active joints actuated by three geared DC motors(20 watt,
gear ratio of 53 : l),respectively. In case of the forelimb, a scapular joint
is mounted to change the width between supporting feet as mentioned in
this research. The hind limb, in particular, is equipped with a miniature
clutch in the knee. It is used to switch between a passive ankle joint and an
active one. The leg spring connected between the foot and the tibia, is used
as the element for storing the energy and reducing the impact on contact
with the ground. Fig. 5 shows the control system layout of AiDIN. Two
embedded controllers and twelve micr~controllers(PICl8f458,Microchip)
are contained in the robot. In case of embedded controllers, the first em-
bedded controller using single board Computer~Pentium-I11800MHz with
compact flash disk, lMbps CAN) and RTLinux is ported as the operat-
ing system. The function of controller is not only control of the walking
posture, but also communication management between micro-controllers
based on CAN(con~rol1erarea network). A voice speech engine and a vi-
sion library(OpenCV, Intel Image Processing Library) are involved in the
other embedded controller(Pentium M. 1.1GHz with wireless network de-
vice). The power of the robot is supplied with a tether cable, although all
248

the other communication is transmitted through wireless LAN(loca1 area

network).

less network supporting

Gyroscope

Rs2321 t TCPLP

I{I 0 Gyroscope module

0 CAN module t
RT-FZF~
I

Micro-Controller

Touch sensor
Force sensor
Clinometer
Current sensor

Fig. 5. System configuration of AiDIN

4. Experiments

The locomotion controller was validated via several preliminary experi-

ments with AiDIN. In the first experiment, self-posture changeability in
AiDIN was tested. The robot was set on a rolling plate as shown in Fig. 6(a).
When the plate suddenly lurched laterally, GLC in AiDIN generated flexion
or extension stimulus to make its posture stable as shown in Fig. 6(b). Even
if a plate is slanted to the right or left, the robot sustained its balance with
GLC.
In the second experiment, a comparison between exceeding and nonex-
ceeding loops in the proposed controller were performed as illustrated in
Fig. 7. In this experiments, the robot walked in a trot gait of 20cmlsec
walking speed. In the first case, the proposed controller was not fully in-
cluded, but only nonexceeding was applied. As shown Fig. S(a), the body
rolls with wide range of rolling angle, even could be unstable during rolling.
 249

In this case, the robot could not walk or fall down at times. When two con-
trol loops were completely included, the robot kept balance with smaller
range of rolling angle as shown in Fig. 8(b). It is noticeable that the robot
can be controlled stably as well as efficiently with the proposed controller.

Fig. 6 . Experiment to verify a self-posture changeability

5. Conclusions
To mimic living creatures has been a contemporary issue in the robotics
researches for long time. It may be questionable to just copy the animal,
but it can be useful if the idea is explained appropriately. In this paper, we
presented a biologically inspired approach on the developing of a quadruped
walking robot controller. Observation on gravity load receptor and stimulus-
reaction mechanism gives us several significant tips on the control of the
quadruped walking robot, which have been analyzed in terms of the robotic
terminologies.
In this paper, we presented the gravity load controller with a rhyth-
mic patterns generator included to control quadruped walking robot. Even
though it is a simple model but it proves, by the experiment results, to
be an effective controlling method in which the sensory feedback is nicely
integrated with the oscillator to produce the rhythmic output signal. It is
also shown in the simulation that this controller gives the robot a potential
ability of adaptation in irregular environments.
250

step 1 step 2 step 3 step 4 Slap 5 Step 6

(b)

Fig. 7. Trot gait on AiDIN

-8
3 -10 ,
m 0 5 10 15 20 25 30 35 40 45 50 55 60
Time [sec]
(a) Trot gait with nonfeedback

9 l:
3 5
2 3
Fo
* -3

m :~~
-10 0 5 10 f5 20 25 30 35 40 45 50 55 60

Time [sec]
(b) Trot gait with feedback

Fig. 8. Body roll angle

Acknowledgment
This research was supported in part by the project of the dual-use tech-
nology for military and civilian missions ( "Development of Quadruped
Robots") of the Ministry of Commerce, Industry and Energy (MOCIE),
KOREA.
 251

References
1. K. Matsuoka, “Mechanisms of Frequency and Pattern Control in the Neural
Rhythm Generators,” Biol. Cybern., vol. 56, pp. 345-353, 1987
2. M. A. Arbib, “The Handbook of Brain Theory and Neural Networks,” A
Bradford Book, The M I T press, 1995
3. H. Kimura and Y. Fukuoka, “Biologically Inspired Adaptive Dynamic Walk-
ing in Outdoor Environment Using a Self-contained Quadruped Robot :
Tekken2” , Proc. IEEE/RSJ Int. Conf. Intell. Robots Syst.,”pp. 986-991, 2004
4. J. Duysens, F. Clarac, and H. Cruse, “Load-Regulating Mechanisms in Giat
and Posture : Comparative Aspects,” The american physiological society, vol.
80, no. 1, pp. 83-133, 2000
5. T. A. McMahon, “Muscles, Reflexes and Locomotion,” Princeton University
Press, 1984
6. L. Robinson, “The Animal Motion Show, Volume 2,” Rhino House,
http://www.rhinohouse.com,2003
7. G. E. Goslow, H. J. Seeherman, C. R. taylor, M. N. McCutchin, and N.
C. Heglund, “Electrical Activity and Relative Length Changes of Dog Limb
Muscles as a Function of Speed and Gait,” J. Exp. Biol., vol. 94, pp. 15-42,
1981
8. T. A. McMahon, “Using body size to understand the structural design of
animals: quadrupedal locomotion,” J. A p p l . Physiol., vol. 39, no. 4, pp. 619-
627, 1975
9. A. V. Hill, “The Dimensions of Animals and Their Muscular dynamics,”
Science Progress, Vol. 38, pp. 209-230, 1950
10. , D’Arcy Wentworth Thompson, “On Growth and Form,” Cambridge Uni-
versity Press, 1917
11. M. Hildebrand, “Analysis of vertebrate structure, 3rd Edition,” New York,
Wiley, 1988
12. Eliot Goldfinger, “Animal Anatomy for Artists,” Oxford University Press
13. S . Hirose, “A Study of design and control of a quadruped walking vehicle,”
Int. J. Robot. Res., vol. 3, no. 2, pp. 113-133, 1984
14. G. Gabrielli and T. H. von Karman, “What Price Speed ?,” Mechanical
Engineering, vol. 72, no. 19, pp. 75-78, 1950
15. D. C. Kar, “Design of Statically Stable Walking Robot: A Review,” J. Robotic
Systems, vol. 20, no. 11, pp. 671-686, 2003
16. J. A. Smith and I. Poulakakis, “Rotary Gallop in the Untethered
Quadrupedal Robot Scout 11,” Proc. IEEE/RSJ Int. Conf. Intell. Robots
Syst., V O ~ .3, pp. 2556-2561, 2004
17. J. P. Schmiedeler, D. W. Marhefka, D. E. Orin, and K. J. Waldron, “A Study
of Quadruped Gallops,” N S F Design, Service and Manufacturing Grantees
and Research Conference, pp. 1-10, 2001
18. T. M. Griffin, R. Kram, S. J. Wickler, and D. F. Hoyt, “Biomechanical and
energetic determinants of the walk-trot transition in horses,’’ J. Exp. Biol.,
V O ~ .207, pp. 4215-4223, 2004
 THE IMPROVEMENT OF STRUCTURAL AND REAL TIME
CONTROL PERFORMANCES FOR MERO MODULAR
WALKING ROBOTS

ION ION*, LUIGE VLADAREANU**, RADU MUNTEANU jr.***, MIHAI

MUNTEANU***
*Department of Technology of Manufacturing, POLITEHNICA University of
Bucharest,ROMANIA, **Institute of Solid Mechanics of Romanian Academy, Department
of Dynamic Systems, Bucharest, ROMANIA, ***Electrical Department of Technical
University Cluj-Napoca, ROMANIA

Starting from the standard MERO walking modular robot, which is a robot on 3x6
degrees of freedom, autonomous, being able to interact with the environment, are
presented the results obtained by the improvement of the structural and real time control
performances. For this purpose kinematics and kinetostatics analysis are performed, and
the mathematic model of the inverted kinematics is determined for controlling the main
trajectory that defines legs’ support and weigh centre positions. Related to this there is
presented an Open Architecture system for the MERO robot position control in Cartesian
coordinates through real time processing of the Jacobean matrix obtained out of the
forward kinematics using the Denevit-Hartenberg method and calculating the Jacobean
inverted matrix for feedback. The obtained results prove a significant reduction of the
execution time for the real time control of MERO robot’s position in Cartesian
coordinates and increased flexibility.

1. Introduction
The improvement of performances concerning the possibility of moving the
walking robots on ground with realistic configuration, in order to increase the
mobility and stability in real conditions, have been realized by developing the
control system in real time of the robot movement trajectory [ 1,2]. Finding the
mathematical model allows for knowing the legs coordinates against robot’s
body and the body position during walking against the support points.
The mechanism within the shft system of MERO modular walking robot
served as an experimental basis [2]. The mechanism is formed by 3 serially
connected active groups of type F2RR. The last two groups form a flat
kinematics chain, functioning in vertical plane. The elements of the first group
are connected to the platform and move in horizontal plane.
This kind of structure imitates quite well the shape of the human leg, is very
simple and reliable, ensuring multiple possibilities of robots’ maneuvers.

252
 253

2. Kinematics Modeling of MERO Robot

Kinematics modeling of the legs' mechanisms related to the main system
attached to the h4ERO modular walking robot is presented in Figure 1. There
are 3 modules with 2 legs each, connected by 3 joint angles. Each leg is
composed by a mechanism with 3 elements connected by 3 rotation actuators.
The Denovit-Hartenberg matrix method in homogenous coordinates is used
in order to determine the legs position against the platform. Following the
analysis of the Hartenberg-Denavit links system for the robot leg mechanism we
have found a quite particular position. This led to decreasing the number of
parameters of the transformation matrix Ai from six to four.

Figure 1. The Hartenberg-Danavit axis system attached to the leg of MERO modular walking robot

The work presents both the transforming connections of the a point's

coordination - from system (i-tl) to the system (i), that are realized by a
transformation matrix Ai - as well as the transforming relations of the Pi point's
coordinations from the system attached to the element (3) of the leg in the
platform system. Considering as already known the legs coordinates h i p , Yoip,
Zoip,the mathematic model with variable parameters of the leg joint angles 8 ,
('j=1,6, i=4,6) have been determined by inverted lunematics analysis. Thus,
knowing the trajectory that the legs should describe, the parameters necessary to
the control system in real time of the robot could be established. By direct
kinematics analysis are calculated XO,~,Yoip,Zoip, and by inverted kinematics
analysis are calculated €).ti, 85i,&. The relative speed is calculated by derivation
of movement equation and the relative accelerations by double derivation.
254

In the general case of the walking robots with 2n legs - the ”myriapod
robots having “n” connected modules - the legs kinematics related to the
environment system have been determined by the same algorithm.
Mathematic modeling of the gravity centre position, which allows the
walking robots to shift on ground with difficult configuration, is presented in the
Figure 2. The geometric centre position 0 is defined as the diagonals
intersection point of the polygon formed by the points where the legs are
connected to the platform, and G(GxG$,fi) as the gravity centre position of robot.
Considering the positions of the walking robot’s legs as Xpi, Ypi, Zpi a
mathematic model has been developed in order to express the kinematics
characteristics of the walking robot’s gravity centre. In order to determine the
position of the supporting polygon against the platform, through the Denevit -
Hartenberg method, where Zi, (i=1,6 or 1,4, j=1,3), and mi, (i=1,6, j=1,3) the leg
mechanism mass, the coordinates of Pi have been transformed from the system
04x4y4z4 into the system Oxoyozo. The coordinates of Pi have been transformed
from the system 04x4y4z4into the system Oxoyozo for to determine the position of
the support polygon against the platform.
The vertical projection of the G system gravity center on the supporting
surface must be inside the supporting polygon in order to obtain stability.

Figure 2. Mathematical modeling of the gravity centre for the MERO modular walking robot

Because the gravity centre positions of each legs mechanism element against the
own system are known, the relations for the gravity robot centre coordinates,
necessary for the stability control of the robot in real time have been established.
..
Knowing the gravity centre position, the speed Xk and the accelerationXk
have been established by derivation.
 255

HYRID POSITION AND FORCE CONTROL OF ROBOTS. As part of

the robot control, the compliance is necessary to avoid power impact forces, to
correct position error of robots or of special mechanical processes devices, and
to allow tolerant relaxation of component elements. The compliance can be
fulfilled either through passive compliance, as in Remote Center of Compliance
(RCC) or through force control active methods [ 51. Passive compliance can
lower robot position capacity. The active compliance can have problems with
sensibility in a rigid environment. Hybrid position and force control of robots
equipped with compliant joints must take into consideration the passive
compliance of the system. The generalized area where a robot works can be
defined in a constraint space with six degrees of freedom (DOF), with position
constrains along the normal force of this area and force constrains along the
tangents. On the basis of these two constrains there is described the general
scheme of hibrid position and force control in Fig.3. For simplification the
coordinate transformations are not noted. Variables Xc and Fc represent the
Carthesian position and the Carthesian force exerted onto the environment.
The selection matrices. Considering Xc and Fc expressed in specific frame of
coordinates, its can be determinate selection matrices S, and Sh which are
diagonal matrices with 0 and 1 diagonal elements, and which satisfy relation:

POSITION CONTROL

I I

Fig.3. General structure of hibrid control.

s, + 'S = I , . (1)
In recent approaches S, and 'S are methodically deduced from kinematics
constrains imposed by the working environment. Let A and B be two matrixes
with full column rank that satisfy the equation A'B = 0 and correspond to the
twist and wrench spaces of constraint, then there can be determined S, and S,
through the relations:
S, = (AtYA)-'A' Y , (2)

S f = (BtY-'B)-'B t Y ' , (3)

where usually Y is symmetrical matrix, positively defined. These theoretical
concepts have been integrated into the open architecture system presented
below.
256

3. Open Architecture System for the MERO Robot Position Control.

As a result of the studies, for the position control of the MERO walking robot
(Figure 4) a real time control system with open architecture OAH has been
conceived. The system has four main components: the programmable logical
controller (PLCO), PC system with open architecture (PC-OPEN),
multiprocessor system PLC (SM-PLC) and the position control system (CPki,
i=1-3, k 1 - 6 ) . The system ensures the implementation of the algorithm for robot
position control in Cartesian coordinates through real time processing of the
Jacobean matrix obtained by direct kinematics of the robot, respectively of the
invert Jacobean matrix for feedback. The method consists in reducing the matrix
J (0) to an upper triangulate form and finding errors in 6 8 joint coordinates
using back-substitution. The joint angle errors 6 8 can be used directly as
control signals for robot actuators. Joint angular errors for actuators on each
robot freedom axis corresponding to the robot’s new desired position generated
in Cartesian coordinates are obtained by processing of Jacobean and inverted
substitution.
The programmable logical controller in decentralized and distributed
structure (PLCO) ensures the control of the freedom axis for the robot execution
elements. The current angle motion values in absolute value Z8,i are transmitted
from PLCO to PC-OPEN. From the PC to the PLC are continuously transmitted
the reference positions on each axis X D according
~ to the trajectory generating
program Xp,. The PLCO is processing 24 analogue inputs, 24 digital output, 18
analogue outputs for robot position and stability control, and other 64 digital
inputdoutputs configurable for auxiliary functions: the hydraulic group control,
motion limitation devices, limitation devices for homing etc. Moreover, the
following analogical and digital input signals are controlled: position
transducers (TPki, i=l-3, k=1-6), proximity transducers, 2xDk horizontal and
2xDk on vertical for orientation, force transducers (Tkp, k=1-6), vertical control
transducers (XM,Ym)for robot stability.
The system PC-OAH (PC-OPEN Architecture) allows the introduction
of new control functions based on supplementary programs. Due to the great
processing speed with operating systems, which allow for programming in
evolved programs, the basic functions can be implemented, using an ExTR real
time multitask executive: interpolation (INP),main program (PP) of the operator
interface, the movement trajectory generating program (MTG), the static
stability control (SSC), robot platform control (RPT), the robot’s walking
shaping program (RWS), compliant function control program (CTRL C) and
the prehensive control program (CTRL P). Supplementary, an interface with
 257

digital camera (IDC) could be installed, for objects recognition and the
communication interface using radio modem, GSM or wireless systems. The
PC-OAH system performs following operations: shapes walking stereotypes,
determines the walking stability limits, ensures the horizontal position of the
platform, maintains the established height of the platform and interprets the
oDerator orders.

Figure 4. Real time control system with open architecture for the MERO modular walking robot

The Multiprocessor System PLC (SM-PLC) and PLCO has the role to
send in real time, through the ARCNET fast communication network, the
angular reference positions for the PIDT position regulator. For feedback, the
current values OKci, (i=1-3, k=1-6), received from the position transducers (TPki,
i=l-3, k=1-6) are transmitted by PLCO through ARCNET. Five processes for
implementing the mathematical model of robot control have been identified [3].
In the active topology for (1) process, each PLC generates an ascendant data
flux from PLCO to PLCS. By calculating the transformed matrix iAj for the leg k
(k=1-6), from the axis i to the axis j, we obtain the coordination matrix in axis j,
finally resulting the coordinates for the robot environment XkC = 'Ak3 =
AI*A2*A3.In the active topology for the process (2), the matrix i-lAkiare stored
for each PLC, the Cartesian coordinates Xki in i axis, by multiplying with lAkj.l,
are determined and the position variation 6Xk, is calculated. The Jacobean
matrix is obtained during the process (3) by an ascending data flow, correlated
258

with process (1). During the matricial calculation ‘Akj,the PLCO is assigned to
J(O,), matrix of (3x1), respectively PLC5 to J(O,). The active topology of
process (4) brings the Jacobean matrix to triangular form by determining the
pivot element and the new matrix Akijelements. The active process ( 5 ) took
place within a sequential process with process (4), determining the value of the
angular error Wi.by inverted substitution.
Position Controllers (CPki,i=1-3, k=1-6), have 18 hydraulic servo-valves
(SVki,k=1-6, i=1-3) as actuators, providing movement control on each free axis,
based upon angular deviations SO“+
The robot’s walking types are issued from 3 block-programs located in
PC-OAH system, namely: the block of walking shaping, which determine the
succession and the way of moving the legs, the block of static stability control,
granting the robot shift so that the system centre of gravity projection may
remain inside the convex polygon formed by the leg’s, the block of platform
control, maintaining the prescribed height and the horizontal position of the
platform.

4. Conclusions and results.

The results of this research have been materialized through MERO modular
walking robot, presented in Figure 5 , whch was built and tested at “Politehnica”
University of Bucharest.
The MERO walking robot’s control system in real time with Open
Architecture ( O M ) ensure flexibility, short time execution, the precision
targets and repeatability of the moving programs, eliminating completely the
closed systems with projects meant for specified applications. Supplementary
developments in order to increase the performances or new functions adding are
possible only by modifying the software relating to the control modules in PC-
OAH for laborious computations, respectively in the PLC multiprocessor for
complex real time control. Besides, addition of new physical modules could
easily be done because of the communication between the programmable logical
controller (PLC) and the input/output modules made by a bus with 3 torsadate
conductors.
Owing to the great computation speed of microprocessor systems and serial
connection links for data transmission, the time necessary for establishing the
inverted matrix is short enough to allow the robot control in real time, with no
influence in performing the other programs. The results show that the time
necessary to perform program for the MERO robot position control in Cartesian
coordinates is 80% shorter.
 259

Figure 5 MERO modular walking robot.

Moreover, the short time execution will ensure a faster feedback, allowing
other programs to be performed in real time as well, like to the prehensive force
control, objects recognition, making it possible that PC-OHA system have a
human flexible and friendly interface.
References
1. Waldron, K.J., (1996) Modeling and Simulation of Human and Walking
Robots Locomotion, Advanced School, Udine, Italy.
2. Ion, I., Simionescu, I., Ungureanu, M., Design and realization of the MERO
modular walking robot, The Eight IFToMM International Symposium on
Theory of Machines and Mechanisms, Bucharest, Romania, Aug. 28, Sept.
1, 2001pp.320-325
3. L.Vladareanu, T. Peterson - New Concepts for the Real Time Control of
Robots by Open Architecture Systems, Machine Building, vo1.55, ISSN
0573-7419 no.l1,2003
4. L.D.Joly, CAndriot, V.Hayward, Mechanical Analogic in Hybrid
PositionRorce Control, IEEE Albuquerque, New Mexico, pg. 835-840,
April 1997
5. Yoshikawa T., Zheng X.Z. - Coordinated Dynamic Hybrid PositionRorce
Control for Multiple Robot Manipulators Handling One Constrained
Object, The International Journal of Robotics Research, Vol. 12, No. 3,
June 1993, pp. 219-230.
This page intentionally left blank
Advances in Walking Robots
This page intentionally left blank
 A BASIC VARIABLES SET BASED SCHEME OF ONLINE
MOTION PLANNING FOR HUMANOID ROBOTS

JIAN WANG, TAO SHENG, HONGXU MA

College of Mechatronic Engineering and Automation, National University of Defense
Technologv, Changsha 41 0073, China.

An online motion planning scheme for humanoid robots is proposed based on basic
variables. The proposed scheme transforms the problem of humanoid motion planning
into the planning of a basic variables set. The chosen basic variables can reflect the
motion state of humanoid robots, including Zero Moment Point (ZMP), Center of
Gravity (COG), Center of Coxa (CoC), foot position and the rotation matrices between
some key frames. Then the planning flow of basic variables are shown and the dynamical
and physical constraints of humanoid motion are taken into account. Finally, the
efficiency and feasibility of proposed scheme are validated by computer simulation.

1. Introduction
Within the world of mobile robots, humanoid robots are of great interest these
years. Originated from biped robots, humanoid robots should be able to pass
obstacles easily and move on uneven terrain optionally, and they can do more
besides walking with whole upper body including functional arms and fingers.
Furthermore, they have huge potential values in entertainment or other fields.
Moving in human-living environment is the basic function of humanoid robots
and motion control plays an important role. Currently, online motion planning
has become an important approach in humanoid motion control.
In Lab of Intelligent Robots, National University of Defence Technology,
biped robots have been studied since 1989. In 2000, the humanoid robot Pioneer
was developed; and “Blackmann” was built in 2003. In order to achieve online
motion planning for “Blackmann”, an online motion planning scheme based on
basic variables set is proposed. Firstly, two kinds of motion planning methods
are reviewed, and requirements of real-time motion control are discussed. Then
the proposed scheme is described, and algorithms of basic variables’ planning
are given. Finally, the effectiveness of proposed scheme is validated.

2. Related Work
Two kinds of usually used motion planning methods are reviewed firstly, and
then the status quo of humanoid online motion planning is investigated.

263
264

2.1. Motion Planning Based on Lumped-Mass Model

A common approach for humanoid motion planning is based on lumped-mass
model. Firstly, the robot is modeled as a point mass, and COG is restricted to
moving on a certain plane. Given these assumptions, the COG motion is
described by two decoupled linear ordinary differential equations of second
degree"]. Then a closed-form solution relying on ZMP is acquired. Generally, in
order to ensure the feasibility of the planning results, a time evolution of the
ZMP is prescribed. The advantages of the method are conceptual simplicity and
the ability to plan motion online. But how to make it effective is difficult.

2.2. Optimization Based Methods

Another method is to solve an optimization problem. Usually most of the DOFs
of the humanoid robot are optimized in joint And the search space for
possible solutions is so large that it is always possible to find a reasonable
trajectory to fulfill the requirements. But the computation efficiency of solution
search is low just because of the large search space of possible trajectories.

2.3. Status Quo of Humanoid Online Motion Planning

Currently, methods used for online motion planning mainly include online
walking motion generation and online mixture and connection of preplanned
motions. LimL3]studied online motion planning using online mixture and
connection of unit gait; Ni~hiwaki[~] proposed an algorithm for online gait
updating based on desired ZMP; Kajita used the ultrasonic sensors to realize
online gait planning of forward walking[51;Afierward, he proposed a method of
3D online walking pattern generationL6].Lorchf7] realized online walking
generation in sagittal plane using technology of stereoscopic vision.
Compared with off-line motion planning, online motion planning cares more
about feasibility and efficiency. And algorithms based on lumped mass model
are chosen instead of time-consumed methods in off-line motion planning.

3. Proposed Scheme
According above analysis, lumped-model based methods are less time-consumed,
but it is difficult to ensure the feasibility of results; optimized based methods can
get good trajectories but cost a long time and it seems unpractical in online
planning. So, lumped-model based methods are appropriate for online planning.
This section describes a basic variable set based scheme of online planning.
Several motion state variables called basic variables are chosen to describe
humanoid motion, the task is to calculate trajectories of basic variables. Various
constraints are considered to ensure the feasibility of planning results.
 265

3.1. Architecture of online motion planning and control

Architecture of online motion planning and control is shown in Fig. 1. The task
of humanoid motion control is divided into three parts: online motion planning,
upper limbs manipulation control and posture/landing control. Here, proposed
scheme mainly deals with online motion planning emphasized in Fig.1.
External sensors

Fig. 1 Architecture of humanoid motion control

3.2. Description of Online Motion Planning

Generally, humanoid motion is described in joint space. But it is difficult to find
a feasible motion sequence in joint space. Therefore, a basic variable set is used
to describe the motion sequence instead of temporal trajectories in joint space.
The basic variable set includes:
xz ,y , ,z, , ZMP trajectory, the criterion of humanoid walking stability;
0 xG,y G ,zG , COG trajectory;
p L F,p R F,positions of both ankle in Cartesian frame;
xc, y c ,zc , CoC in Cartesian space. CoC is the origin of the frame attached
to coxa, and coxa is the part which connects both hips.
8,, ,efl, 8,, - ighr ,rotational DoFs of both hips, variables for turning motion.
FspRE S0(3), rotation matrix of frame attached to coxa relative to frame
Coxa

attached to supporting foot;

E S0(3), rotation matrix of frame attached to supporting foot relative
to inertial frame;
: S R E S 0 ( 3 ) , rotation matrix of frame attached to swing foot relative to
frame attached to coxa.

3.3. Basic Variables Planning

Since the motion sequence is described using basic variables, the planning
methods of basic variables are proposed as forward walking for example.

3.3.1. Humanoid Dynamic Model

Generally, relationship between Z M P and COG is used to describe the dynamics
of humanoid robots, which is shown as follows:
266

xz =x, -(zc -zz).X,/g (1)

Y Z = Y C -('G -ZZ)'YG/g (2)
Here, [x, y G z,] and [xz y z z Z ] are COG and ZMP trajectory in
inertial fiame respectively, g is gravity.

3.3.2, ZMP and COGPlanning

This section mainly discusses the planning of ZMP and COG.Here, three points
are considered: firstly, the model described in (1) and (2) is used; secondly, the
idea of boundary relaxation proposed by T. Sugihara[*]is adopted; thirdly, spline
functions are used as the form of COGtrajectory.
In one walking period T , given desired boundary values of ZMP and COG
in sagittal plane:
dX=[dXZ(0) dxz(T) dXG(T) dx,(T)r (3)
Given ZMP trajectory in sagittal ( x -axis) plane:
(4
xz = f Z M P (Po tY xz ( 4 2 xz (r))
3 (4)
Here, f z M pis using the form of cubic polynomial, and p o is the parameters
set described by ZMP boundary conditions and walking speed. Using equation
(l), the solution is drawn up using spline function:
G
' ( t ) = fCoG ',
'2 (O), x Z (5)
Here, p is unknown and boundary conditions of ZMP and COG are used to
solve p Similar to [S]. Let x = [xz (0) xz (I") xG ( T ) x, (T)]' , and set t = 0

g(p0 p , t , x 9 xG (O), xG
9 = (6)
Constraint inequalities are formed using walking stability condition:
&z (O), xz ( T I )< 0 (7)
Solving (8) will obtain appropriate boundary conditions under ( 6 ) and (7):
(x-'x)T ~ ( x - d x ) /2 -+ min (8)
-
Here, A = diag{a,} , (ai > 0,i = 1 4) . Using idea of penalty hnction,
problem (8) can be replaced by (9):
(x-~x)T A ( x - ~ x ) 2/ + i
C, (max{h, ,0})2 +
J
D j gj z + min (9)

Where C,, D j are weight parameters, and C,,D j > 0 . Using calculated
boundary conditions to compute p ,ZMP and COGtrajectories are obtained.
 267

3.3.3. Footprint Planning

In forward walking, only the sagittal (in x - axis ) and vertical (in z - axis ) parts
are considered, and only the positions of ankle point are planned.
In humanoid walking, single and double support phases arise in sequence.
Positions of supporting and swing foot in t = O are (xsp(0),zsp(O))and
(xsw (0),zsw (0)). xsw ( T ) = xsw (2") + 2 . S is decided by pace S . When ZMP
enters into the coming single support region, the swing leg begins to rise. When
t = T , the swing foot reaches the desired position xsw ( T ) and xsw ( T )= 0 . The
time to start foot lifting T, is decided by planned ZMP trajectory in simulation,
and it is calculated using flow chart in Fig. 2, where R is the target single
support region. Afterward, the sagittal and vertical parts of temporal trajectory of
swing foot can be planned......................................................................................................................
/for i=l:N
i if (xz(i.TIN),y,(i.TIN))Ef
! then T, = i . T I N ; b r e a k
i
i else i = i + l
i
j end
i.....................................................................................................................
end
Fig. 2 Time Decision of Swing Foot Lifting

3.3.4. Calculation of CoC trajectoly

The positional relationship between COGand CoC is described as follows:

pG = ~ m i p i f ~ = p , + ~ m i ~ R C p i ~ M (10)
i=l i=l

Where, p c = (xG,y G,zG ) ,p , = (xc ,y , ,z , ) are positions of COG and CoC

in inertial frame, :R E SO(3) is rotation matrix of frame attached to coxa
relative to inertial frame, ' p i = ( c n i , c y i , c z i )is position of link i in frame
attached to coxa, mi is the mass of link i ,and M is the mass of the robot.
Firstly, the mappings from joint angles to each link's center of mass are built
and :R is calculated using E R and Fs:R . Generally Fs:R can be obtained
beforehand, and E R is decided by configuration of supporting leg.
,OR(&4)=F$R.ER(Q, 4') (1 1)
Pi = f F K - i (12)
. .
9'( 'i

1
Here B = (0,,-..,Bj ,...On is the joint variables set, and 4 4' ci are
kinematical parameters sets. Then (10) can be rewritten using (1 1) and (12):

PG = PC +Ami
i=l
,OR(', 4yfF K -i ('7'i )lM (13)

If initial configuration of the robot is ascertained, p c ( 0 ) , B(0) p G ( 0 )

are known. Set AB(0) = 80)- B(0) , owing to AB + 0 , p , (1) can be formed:
268

Similarly, p c (k + 1) can be calculated utilizing p G(k + 1) and results of the

k times inverse kinematics as shown in Fig.3:

3.3.5. Inverse Kinematics

After positions of CoC and both ankles acquired, inverse lunematics is under
way using transformation matrix, Jacobian and rotation matrices E R , Fs:R
and : Q R , as seen in Fig.3. Either leg of the humanoid robots can be regarded
as a series configuration with 3 links and 6 DOFs. Firstly, the transition matrices
are computed using D-H coordinates, then coxa Jacobian JsuPp,, between joint
angles of support leg and the coxa configuration and swing foot Jacobian J,,
between joint angles of swing leg and the swing foot configuration are derived.
Finally, RMRC [’I method is used to compute the joints angles.

3.3.6. Considered Constraints

During motion planning, two kinds of restrictions are considered, including
stability, friction constraints and physical constraints. Stability conditions are
considered in ZMP and COG planning, so as to ensure the feasibility of walking;
while in high-speed walking, friction constraints should be counted. Physical
constraints considered in inverse hnematics include leg length restriction, joint
angle and velocity restrictions. Leg length restriction means the distance between
ankle and hip should be less than leg length. Joint angle and velocity restrictions
mean the planned joint angle and angular velocity should be in a reasonable span.

4. Simulation and Analysis

Taking humanoid robot “Blackmann” as the object, forward walking simulation
in corresponding platform is used to test proposed scheme.
It is supposed that pace S = 0.2~1,walking speed v = 0.2mls , step period
- FSP R
T = 1s ,zz = 0,zG = const , O r 0 , - , ~- 0, Bror-righl= 0 and Coxa ,Fs:R ,ZR are
identity matrices, and the physical parameters of “Blackmann” shown in Tab.1
 269

are used in simulation. Using aforementioned methods, Z M P , COG and CoC

trajectories as well as joint angles are obtained (seen in Fig.4 and Fig. 5). And
kinematics simulation is carried out as seen in Fig.6. Then through dynamic
simulation, it is shown that the recalculated Z M P (seen in Fig.4) still locate in
stable margin, so results using proposed scheme are feasible in humanoid
walking. In simulation environment (Pentium4 2G, 5 12M RAM,Windows X P ) ,
one step planning including inverse kinematics costs about 450 ms with 100
pointshtep. While in previous off-line motion planning [Io1, calculation of one
walking cycle will cost tens of seconds. So, proposed scheme can be used online.
Table 1. Physical Parameters of “Blackmann”

Mass(Kg) Length(mm) Width(mm) Height(mm)

Total 63.5 \ \ 1545
Head 0.5 \ \ 200
Trunk 31.82 \ \ 580
Arm 1.40 \ \ \
Thigh 7.06 \ \ 350
Shank 3.94 \ \ 340
Foot 3.16 \ \ 95
Sole \ 270 175 \

Fig.4 Planned ZMP, COG,CoC, Footprint Fig.5 Joint Angles of Right Leg
Trajectories and Recalculated ZMP

.
Fig 6 Snapshots of Walking Sequence in Simulation

5. Conclusion
An online motion planning scheme based on basic variables set for humanoid
robots is presented. Proposed scheme transforms humanoid motion planning
problem into basic variables’ planning. ZMP and COG trajectory are obtained
using boundary conditions. Then positional relationship between COG and CoC,
270

is used to calculate CoC trajectory online. Finally, joint angles are calculated by
inverse kinematics. Walking experiment using proposed scheme is not finished
up to now and it can’t cope with real-time motion control alone in unstructured
environment, so ongoing and future work includes experimental evaluation of
proposed scheme using “Blackmann”.

Acknowledgments
This research is supported by grant 60475035 of the Chinese National Natural
Science Foundation.

References
1. S. Kajita, F. Kanehiro, K. Kaneko, K. Fujiwara, K. Harada, K. Yokoi, and
H. Hirukawa. “Biped walking pattern generation by using preview control of
zero-moment point”, In Proc of the IEEE ICRA, Taipei, Taiwan, sep. 2003.
2. G. Bessonnet, S. Chesse, and P. Sardain. “Optimal gait synthesis of a seven-
link planar biped”. The Intl Journal of Robotics Research, vo1.23, no.10-11,
pp. 1059-1073,2004.
3. H.-0. Lim, Y. Kaneshima, and A. Takanishi. “Online W a k n g Pattern
Generation for Biped Humanoid Robot with Trunk”. Proc. of IEEE ICRA,
pp. 3 111-3 116,2002.
4. K. Nishiwaki, S. Kagami, Y. Kuniyoshi, M. Inaba, and H. Inoue. “Online
Generation of Humanoid Walking Motion based on a Fast Generation
Method of Motion Pattern that Follows Desired ZMP”. Proc. of IEEERSJ
Int. Conf. on Intelligent Robots and Systems, pp. 2684-2689, 2002.
5 . S. Kajita and K. Tani. “Adaptive gait control of a biped robot based on real-
time sensing of the ground”. Proc. of IEEE ICRA, Lausanne, Switzerland,
1996. pp. 570-577.
6. Shuuji Kajita, Fumio Kanehiro, and etc. “A Realtime Pattern Generator for
Biped Walking”. Proc. Of IEEE ICRA, 2002.
7. 0. Lorch, J. F. Seara, K. H. Strobl, U. D. Hanebeck, and G. Schnudt.
“Perception Errors in Vision Guided Walking: Analysis, Modeling, and
Filtering”. Proc. of IEEE ICRA, pp. 2048-2053,2002.
8. Tomomichi Sugihara and Yoshihiko Nakamura. “A Fast Online Gait
Planning with Boundary Condition Relaxation for Humanoid Robots”. Proc.
of the 2005 IEEE ICRA, Barcelona, Spain, April 2005.
9. Deepak Tolani, Ambarish Goswami, and Norman I. Badler. “Real-Time
Inverse Kinematics Techniques for Anthropomorphic Limbs”. Graphical
Models, vo1.62,pp.353-388,2000.
10. Jian Wang, Tao Sheng, Hongxu Ma and Haili Qin. “Design and Dynamic
Walking Control of Humanoid Robot Blackmann”, Proc. of the 6th World
Congress on Intelligent Control and Automation, Dalian, China, pp.8848-
8852,2006.
 A HOPPING MOBILITY CONCEPT FOR A ROUGH TERRAIN
SEARCH AND RESCUE ROBOT

SAMUEL KESNER
JEAN-SEBASTIEN PLANTE
STEVEN DUBOWSKY
Mech. Eng. Dept., Massachusetts Institute of Technology, 77 Massachusetts Ave.
Cambridge, MA 02139, USA

PENELOPE BOSTON
Earth & Env. Sc. Dept., New Mexico Inst. of Mining and Technology, 801 Leroy Place
Socorro, NM87801, USA

A new search and rescue concept based on the deployment of teams of small spherical
mobile robots (“Microbots”) has been proposed. In this concept, hundreds to thousands of
cm-scale, sub-kilogram Microbots are released over a search site such as collapsed
building rubble or caves. Microbots use hopping, bouncing, and rolling to infiltrate
subterranean spaces in search of possible survivors. Key technologies enabling Microhots
are the use of high energy-density micro fuel cells combined with low cost and
lightweight dielectric elastomer actuators. The paper presents recent work demonstrating
the feasibility of Microbots mobility in rough terrain. Experimental studies have
demonstrated the possibility of using dielectric elastomer actuators to generate
autonomous hops. High efficiency hydrogen fuel cells have also been used to power
dielectric elastomer actuators. Simulation results show that Microbots of proper diameter
and hop height can successfully traverse very rough terrains. These results suggest that
teams of Microbots can effectively be used for search and rescue missions.

1. Introduction
Events such as the 2005 Pakistan earthquake and the 2001 September 11
terrorist attacks demonstrate the need for new effective search methods in rough
terrain, see Figure 1. Current search methods for rough terrains are limited.
Remote imaging techniques to identify subterranean features, including ground
penetrating radar, ultrasonic imaging, and resistive imaging, have been
developed [ 1,2]. However, these methods are limited in resolution and depth
due to soil properties. They also cannot detect the presence of disaster survivors
in difficult to reach locations. The “dog and pole” method is still the best
civilian search technique.

271
272

(8) (b)
Figure 1. Typical search and rescue sites: (a) 2005 Pakistan Earthquake (b), September 1I, 2001 I

A new approach for search and rescue in rough terrains based on hopping
robots, called Microbots has been proposed [3]. As shown in Figure 2(a),
Microbots are small spherical robots of about 10 cm in diameter. The search and
rescue approach consists o f deploying hundreds or thousands o f Microbots over
a search site. The Microbots use hopping, bouncing, and rolling to navigate
rough terrains in search of survivors. Due to their small size, Microbots can
difkse inside rubble cavities to find internal passage leading to protected

Figure 2. The Microbot concept: (a) artist representation, (b) progression in rubble.
Microbots are powered by high energy density Proton Exchange Membrane
(PEM) he1 cells to assure long lasting energy supply. The mobility system is
actuated by lightweight and low cost Dielectric Elastomer Actuators. Microbots
are equipped with onboard miniature sensors such as cameras and chemical
“sniffers” to tract and identify survivors. Their communication systems relay
information between each other and a command center. Microbots components
are protected by a strong plastic shell that absorbs shocks.
 273

Microbots missions differ from conventional robotic missions that often

use a single highly capable agent. Instead, Microbots missions use a very large
number of low cost and simple agents, bringing a high degree of redundancy
and robustness. Individual agent losses are acceptable without failing the
mission objectives. Also, the low costs of Microbots make them disposable
which eliminate the need for post mission recovery.
The mobility of Microbots in rough terrain is one of several important
technical challenges that must be carehlly understood before Microbots become
a reality. Hopping robots have been proposed for space exploration and
reconnaissance applications [4,5,6]. Most of this work focuses on the
development of hopping mechanisms for relatively heavy robots (>lkg) and are
not appropriate for lightweight Microbots. Developing a practical mobility
system for small and lightweight hopping robots, especially for rough terrain
environments, has not been addressed.
This paper studies the feasibility of the Microbot mobility concept for
search and rescue missions using experimental validations and simulations.
An experimental Microbot prototype powered by Dielectric Elastomer
Actuators has been constructed. It achieved hops of 38 cm with actuators that
have less than one-half the thrust of the Microbot reference design, due to
current laboratory fabrication limitations. Methods to build more powerful
actuators are currently being developed. This result shows the technology to be
suitable for Microbots.
Experimental miniature PEM fuel cells using hydrogen have been used to
power Dielectric Elastomer Actuators. Conversion efficiencies have been
measured across the energy chain and projected Microbot performance are
reported here. These experiments show the concept is viable for 1000 hops
missions.
Simulations of the Microbot mobility show the effect of Microbot diameter
and hop height on travel distance in rough terrain. The simulations shows that a
Microbot diameter of 10 cm with a projected hop height of 1 m give reasonable
rough terrain mobility.
The general conclusion of this paper is that, assuming reasonable
technology progress, Microbots could effectively move in rough terrains for
search and rescue missions.

2. Microbot Mobility Concept

The mobility mechanism concept is illustrated schematically in Figure 3. Energy
is stored in the form of hydrogen gas in a metal hydride storage vessel.
274

Hydrogen reacts with atmospheric oxygen in the PEM fuel cells to generate
electricity. Pure oxygen could be stored onboard for anaerobic applications. A
small lithium polymer battery is used to level power consumption peaks. The
DEA pumps mechanical energy into a bi-stable spring over one or more
actuation cycles. When a predefined energy level stored in the spring is reached,
the energy is released to provide hopping power.

Fuel Cell 4

Li-Po

spring

Figure 3. Schematic of the Microbot mobility concept on rough terrain.

Microbots are self-righting so that after each hop, they return to an upright
position. Directionality can be provided by number of mechanisms, including
small additional DEAs that tilt the Microbot prior to hopping. Directionality
consumes little energy compared to hopping and is of secondary importance at
this stage in the Microbot development.
The Microbot mission concept exploits the high force-to-weight and
simplicity of DEAs [7,8,9,10]. These qualities make DEAs very attractive for
Microbot missions since a large number of strong and lightweight actuators are
needed. Another application exploiting the same characteristics of DEAs is
binary actuation [ 113. DEAs are also low power / high energy density devices
that match well with the proposed fuel cell energy storage technology.
The preliminary design specifications of the mobility system for search and
rescue missions are summarized in Table 1. These numbers are referenced
throughout this paper.
Table 1. Microbot Mobility System Specifications.
Parameter Values
Microbot Diameter 10 cm
Hop Height lm
Microbot Mass 100 grams
Min. Autonomy 1000 hops
Min. Hop Frequency 2 hops / minute
 275

3. Dielectric Elastomer Actuators Powered Prototype

A simplified Microbot prototype has been built to demonstrate the feasibility of
using DEAs to make a Microbot hop with an onboard energy source. The
prototype is shown in Figure 4. A conical shaped DEA pumps energy into a pair
of power springs. When a prescribed number of pumping cycles is reached, the
stored mechanical energy is released and the Microbot hops. The transmission
structure is made from carbon fiber. The power springs are made of carbon fiber
strips. The strips are normally flat and are mounted in a buckled state. The
combined mass of the actuator, transmission, and power springs is 18 grams.
The energy source for the prototype is a single 145 mAh Lithium-Polymer
cell that weighs 5 grams. A custom electronic circuit using a pair of EMCO Inc.
miniature DC/DC converters generates the 8.8 kV needed by the DEA. The
Microbot prototype is shown in Figure 5. The mobility system and electronics
are enclosed in a 10.5 cm diameter PETG shell. The 46 grams Microbot reaches
vertical hop heights of 38 cm. Each hop requires 35 actuator pumps.

Power Springs 8

btgure 4 Mobility system prototype

I

38 cm

Ftpllrc 5 Autonomous Microbot prototype perfomitng hops of 38 cm

276

The Microbot prototype clearly indicates that DEAs can power lightweight
hopping robots. The total mass, hop height, and pumping times of this hand-
fabricated prototype are within reach of the target values of Table 1. Achieving
the specifications of Table 1 appears possible with improved manufacturing
techniques and further design optimization.

4. Fuel Cells Energy System

A hydrogen fuel cell energy system is proposed to power Microbots. For
hopping robots, energy consumption during hopping is proportional to the
system weight and hop height. An analysis of a fuel cell energy system's
requirements is therefore related to the system mass via the hop height
requirement and the fact that the hydrogen fuel and associated storage device
has a non-negligible mass. The performance of the energy system is
characterized by the relationship between the number of hops and Microbot
system mass:

where Nhops is the number of hops, qT is the total energy conversion and
hopping efficiencies, qreg is the low voltage regulator efficiency, Efc is the
electric total energy generated by the fuel cell system through the conversion of
hydrogen, Perec is the power consumption of the onboard electronics, m is the
mass of the Microbot, g is the acceleration of gravity, and h is a hop height of
1 m, and t is the length of the mission. In this case t is assumed to be 3.5 days.
An experimental fuel cell power system was constructed and used to power
a Microbot prototype, see Figure 6. Values for the fuel cell efficiency, low
voltage regulation efficiency, and high voltage conversion efficiency were
found experimentally to be approximately 70%, 90%, and 30% respectively. A
detailed explanation of the analysis has been presented [12]. Figure 7 shows a
plot of the hops/mass relationship. The target of 1000 hops can be reached with
a Microbot mass of about 100 grams.

5. Mobility Simulations
The experimental demonstrations conducted to date indicate that the Microbot
specifications of Table 1 are realistic. Simulations studying the effect of key
parameters such as Microbot diameter and hop height on performance have been
performed.
 277

Low Voltage

Electronics

Figure 6 The hydrogen fuel cell experimentalsetup.

50 100 150 200

Microbot Mass (9)
Figure 7: Number of Microbot hops as a function of system mass.

5.1. ~ i m u ~ a t Approach
~on
A tunnel with debris was selected to represent a disaster area, such as a
collapsed passage in a mine, a subway tunnel after an earthquake, or the interior
of a collapsed building. The simulated terrain was generated in Solidworks
CAD software as an assembly of individual solid bodies, see Figure 8. The rock
pile is composed of 300 rocks of different sizes randomly grouped together into
a pile approximately 5x4x0.85 m. The tunnel diameter is 5 m and its length is 60
m.
The simulations are conducted with MSC So€tware’s ADAMS dynamic
simulation software. ADAMS allows the definition of mass properties, body
forces, and body interaction constraints and forces. The directionality of the hop
is controlled by a Simulink ~ a t h w o r k smodel
) c o ~ u n i c a t i n with
g ADAMS.
The Microbot interacts with the environment through hopping, bouncing
and rolling on the terrain. Hopping was modeled as an impulse force between
the Microbot and the terrain. The hopping direction depends on the angle that
the impulse is applied relative to the Microbot’s body.
278

Figure 8. A Microbot traversing the simulated terrain.

The bouncing and rolling are modeled as an impact contact model with
fiction. The model generates a variable force between the Microbot and the
terrain in a direction that resists the relative motion of the two bodies. The
impact force is modeled as a nonlinear spring/damper system:

where k is the spring stiffness constant, b is the position dependant damping

coefficient, and Ax and Ak are the relative displacement and velocity.
The friction force used in the simulations is standard Coulombic friction
with a velocity dependant friction coefficient, p(v). The parameters used in the
simulations were estimated from laboratory experiments in which the behavior
of a Microbot on compacted dry sand and rocks was observed. Table 2
summarizes these values.
Table 2: Values used in the impact contact model.
Parameter Value
k 240,000 N/m
b (sand) 10 N-s/m
b (rock) 0.5 N-s/m
static 2
p dynamic 0.15
Stiction Transition Velocity 0.01 m i s
Friction Transition Veiocity 0.1 mis

5.2. Results and Discussion

A large number of simulations were run.Microbot diameters of 5 , 10, and
20 cm and hop heights of 50, 100, 150, and 200 cm were used. Each
 279

combination of hop height and size was simulated with a different starting
positions spread over an approximately 2 meter area. The Microbot mass is
fixed to 100 grams. In each simulation, the Microbot starts approximately 2 m
from the rock pile and has 14 hops to overcome the obstacle.
Figure 9(a) shows the rate of successful trials as a function of Microbot hop
height. Success is defined as completely overcoming the rock pile. Failed trials
were caused by three failure modes: 1) entrapment, when a Microbot is trapped
by a group of rocks and is unable to hop out, 2 ) low hop height, when a
Microbot is caught because it is unable to hop over a rock, and 3) bouncing
away is when the Microbot hops in such a way that it bounces away from the
rock pile and must start again and can not complete the task in 14 hops. The
consequence of this is not a failure per se but an undesirable delay. Figure 9@)
illustrates the most common failure modes for the Microbots as a function of
hop height and Microbot diameter.
Success Rate
16
14
8o.aox
8o.aax 3 12
70.W g 10
2 eonax 0 8
UI 5 a . o ~
4a.om -.-
D
0 6
30.00%
20.00%
24
2
10.00%
o.aa% a
50 1w 150 200 iop Hsigl

Hop Height (cm) Microbot Dia. 5cm 10cm

Figure 9.Simulation results: (a) The rate of successful attempts as a function of hop height, and (b),
the failure modes as a function of Microbot diameter and hop height.
The results show that all trials with low hop height resulted in failure. This
suggest that a hopping robot can overcome a complex obstacle only if the hop
height is greater than a characteristic height of the features on which it climbs,
in this case approximately 0.85 m. Here, a hop height of 1 m leads to some
success. Hence, hop height should be maximized. However, increased hop
height trades off with larger power consumption and mechanism weight.
The results also indicate that small Microbot size result in greater
entrapment. The rock pile was randomly assembled and is not an exact model of
a real pile of rubble found in disaster zones. However, it can be deduced that the
maximum size Microbot should be select to minimize the chance of entrapment
while still being able to fit inside the smallest openings it may need to pass
through. The bouncing away failures seen in the simulations are not as much of
280

a concern since they only retard the Microbot’s progression. These failures
could be improved or eliminated by effective path planning.

6. Conclusion
This paper analyzed the feasibility of the Microbot mobility system in rough
terrain. An autonomous hopping DEA prototype has performed 38 cm hops in
the lab. A fuel cell power system experiment and analysis indicates that a 100
grams Microbot could perform about 1000 hops. Simulations suggest that a 10
cm diameter Microbot performing hops of 1 m high could succeed in rough
terrain typical of search and rescue sites. These results confirm that, with
reasonable technology development, the Microbot system could become an
effective tool for search and rescue missions.

References

1. A. Chamberlain, W. Sellers, C. Proctor, and R. Coard, “Cave Detection in Limestone using

Ground Penetrating Radar,” Journal ofArchaeological Science 21, 957-964 (2000).
2. W. Sellers, and A. Chamberlain, “Ultrasonic cave mapping,” Journal of the Cave Research
Electronics Group 28, 18-19 (1997).
3. S. Dubowsky, JS. Plante, and P. Boston, “Low Cost Micro Exploration Robots for Search and
Rescue in Rough Terrain”, IEEE International Workshop on Safety, Security and Rescue
Robotics, (2006).
4. P. Fiorini, S. Hayati, M. Heverly, and J. Gensler, “A Hopping Robot for Planetary Exploration,”
in Proc. ofIEEE Aerospace ConJ, Snowmass, CO, 1999.
5. S. A. Stoeter, P. E. Rybski, M. Gini, and N. Papanikolopoulos, ”Autonomous stair-hopping with
scout robots,“ in IEEE/RSJ International Conference on Intelligent Robots andsystems,
Lausanne, Switzerland, 2002, pp. 721-726.
6. G. J. Fischer and B. Spletzer, “Long range hopping mobility platform,” in SPIE Unmanned
Ground Vehicle Technology Conference, Orlando, FL, United States, 2003, pp. 83-92.
7. R. Kornbluh, R. Pelrine, Q. Pei, S. Oh, and J. Joseph, “Ultrahigh Strain Response of Field-
Actuated Elastomeric Polymers,” Proc SPIE Smart Structures and Materials 2000 (EAPAD)
3987,5 1-64 (2000).
8. R. Pelrine, R. Sommer-Larsen, R. Kornbluh, R. Heydt, G. Kofod, Q. Pei, and P. Gravesen,
“Applications of Dielectric Elastomer Actuators,” Proc. SPIE Smart Structures and Materials
2001 (EAPAD) 4329,335-349 (2001).
9. A. Wingert, M.D. Lichter, S. Dubowsky, and M. Hafez, “Hyper-Redundant Robot Manipulators
Actuated by Optimized Binary Dielectric Polymers,” Proc. SPIE Smart Structures and Materials
2002 (EAPAD) 4695,415-423 (2002).
10. JS. Plante, and S . Dubowsky, Smart Materials and Structures 16, S2274236, (2007).
11. JS. Plante, L. Devita, and S. Dubowsky, “A Road to Practical Dielectric Elastomer Actuators
Based Robotics and Mechatronics: Discrete Actuation,” Proc SPIE Smart Structures and
Materials 2007 (EAPAD), (2007).
12. S . Kesner, JS. Plante, P. Boston, T. Fabian, and S. Dubowsky, “Mobility and Power Feasibility
of a Microbot Team System for Extraterrestrial Cave Exploration,” Proc. of IEEE Robotics and
Automation Conf, Roma, Italy, 2007.
 A PROPOSAL FOR BIPEDAL LOCOMOTION USING
GYROSCOPIC EFFECT*

PULKIT KAF'UR'
Department of Mechanical Engineering, Punjab Engineering College,
Chandigarh, INDIA

RAHUL MUKHI
Department of Production Engineering Punjab Engineering College,
Chandigarh, INDIA..

VINAYAK+
Project Assistant, Centre for Product Design and Development,
Indian Institute of Science, Bangalore, INDIA.

In literature, there have been bipeds which involve the use of a large number of actuators
and those which use gravitational force to propel i t . ; passive dynamic walkers. Study
has also been canied out in context to connect the similarities between gaits of two-
legged robots and their biological counterparts. The present work focuses on the idea of
joining these two extremities with a novel approach to actuate two-lggged systems. This
paper proposes the idea on how to use gyroscopic effect, as means of propulsion, for
bipedal locomotion. By modeling and simulating a two-legged system in ADAMS, the
authors prove the feasibility of the proposal. The biped illustrated in the paper consists of
a rotating flywheel which provides gyroscopic couple in order to move the system
forward as the legs march in the frontal plane of the system. Preliminary results regarding
the effect of the mass and speed of the gyroscope have been shown and the proposal is
also supported by a mathematical analysis which provides a measure of the leg saoke in
terms of the angle rotated by the system about the leg in support phase. The present work
does not consider the start and the stop positions of the system but aims primarily at
proving the concept of how gyroscopes can be used to propel a bipedal system.

1. Introduction:
Bipedal locomotion involved the movement of two legged walking systems. Due
to anthropomorphic similarity of bipedal gait, and potential scope of
development of gait assistive devices, there has been a lot of interest in this field.

* Corresponding Author:vinavak.oro@gtnail.com
+ ~maii:virtua~puru@~mail.cotn

28 1
282

Research in Bipedal locomotion had previously focused on achieving the

desired gait with over actuated systems, and advanced control algorithms,[3] etc.
In 1990, Tad McGeer [ 11 showed that a simple planar mechanism with two legs
could be made to walk stably down a slight slope with no other energy input or
control, using the inherent dynamics of the system. This system acts like two
coupled pendulums and given sufficient mass at the hip, the system will have a
nominal trajectory that repeats itself. Ruina eta1 [6] made a 2d passive walker
with knees, which could walk on level ground.
The aim of the present work is to propose a means to propel a two-legged
(knee less) system on level ground using gyroscopic effect. Next, we verify
whether the system that has been proposed is feasible at its most fundamental
level. This is done by modeling the system in ADAMS [7] and simulating it
under certain conditions and seeing whether the system works. In case the
proposed system works, the third aim is to investigate certain important
parameters with respect to other parameters in context to how they vary during
the work period of the system.

2. Gyroscopic propelled locomotion Concept

The use of Gyroscopes in mechanical systems is not new. However, in past they
have been used for providing stability, and steering capabilities. Gyrover [3]
developed at Carnegie Mellon University, has an internal mechanical gyroscope
that allows it to stabilize and maneuver over rough terrain, slopes etc. Also
gyroscopes have been used for providing orientation estimate of a mobile robot
with respect to his environment [2], for mobile hopping robots stabilized by
rotating gyroscope [5]. The motion of biped is based on the mechanical principle
of gyroscopic effect. A rotating flywheel, tends to presses, (rotate about an axis),
at right angle to applied torque, according to fundamental equation of gyroscopic
effect:

T =I, XWXW,

Where, T is the torque applied, I, is the moment of inertia of flywheel, and o is

its angular velocity, a, is the angular velocity of precession. We aim to utilize
this gyroscopic couple for propelling a bipedal system (knee less) forward on
level ground. This is achieved by mounting a rotating flywheel at the hip and
actuating one of the legs vertically up. Then due to gyroscopic effect, the swing
leg, presses about the stance leg axis, and the biped is propelled forward. A
control algorithm is then developed, for controlling the on time of the gyroscope,
for desired stride length.
 283

3. ~ethodology:

I
Fig.2: Top View of the trajectory of biped

3.1. Assumpt~ons:
The model of the biped was constructed in ADAMS, to see if the conceptual
model worked under certain conditions. The assumptions made in the model
were that, the stance leg of the biped is considered fixed on the ground, for the
purpose of simulation. Secondly, for our analysis we have considered the
intermediate position of the gait, when it moves from an angle -8 to +0 as
illustrated in Fig. 2. Also all the links and moving parts have been considered
rigid in the model.

3.2. The Control Strategy:

The gyroscope cannot be stopped immediately, as the biped might overturn due
to the inherent dynamics of the system. Hence a control strategy is needed to
determine the on-time of the gyroscope. This strategy must follow a closed
recursive loop, as all the energy must be returned back to the system. The
following mathematical analysis provides a logical direction to our choice
The variables involved in the control algorithm are as follows:
1. I, - Moment of inertia of gyroscope
2. I, - Moment of inertia of body
284

3. Wg- Angular velocity of gyroscope

4. W , - Angular velocity of precession

5. L - Distance between supports
6. d - Distance between leg axis and support

I Front View

fl'-%K

Fig. 3. Stick models showing the intermediate motion of Biped

When Biped starts from rest to move angle 8

When leg is lifted up couple acts due to gyroscopic effect,

C=I,XW,XW,
(2)
I , xu, XW,
Force on supports, F= (3)
L
Torque on body about leg2 is given by

Tbody= F x { ( L + d ) - d }
(4)
a T b o d y= F X L = I , XU, X U p

I , xu, XW,
Torque can also be represented as I X a , thus: = 1 (5)
l b
Now let body be turned about leg2 by an angular displacement p . The time
taken for this amount of angular displacement is given by:
 285

t, =(?) 112

And hence the angular velocity a,, attained after time ‘t,’ will be:

lot1= a x t , = (2ap)1’2 (7)

The further angular displacement of legl is dependent on the time taken to fall
on the ground. If assumed that legl will fall on the ground with same angular
velocity U p, then the angular displacement is given by: =

The total angular displacement therefore is given by p + 2,8 = 3p which

further is equal to 6 .

Intermediate Position from - 8 to + 8

For total angular displacement of 28, 3,8, = 26
2
(9)

I , x w, x w p
a=
I,
1
1
Time for this angular displacement, p, = - X a x t 2
2

I I , xw, x w
PI =2 t2
I,
For time t the motors should be actuated to lift the leg and same to drop the leg
to the ground after gyro is turned off. Total time taken to move 28 is 2t.

The preceding mathematical analysis provides a basic framework, which guides

286

our logic towards the working of the system. Also, the gyroscope should not be
suddenly stopped or started. This is because the biped might turn over, if the
gyroscope is suddenly started or stopped, in order to conserve the angular
momentum of the system. The rate of stopping the gyroscope is yet to be
determined. Complete dynamic analysis and equations of motion need to be
derived for that, This rate may also be calibrated from experimental study of a
working model, in near future.

4. Results and Issues involved:

The model of the biped was simulated in ADAMS, and gyroscopic propulsion of
the bipedal system was observed. The system was simulated for its motion under
a set of parameters and the variation in certain significant variables was noted for
the variation of the following parameters:
Mass of Gyroscope
Speed of gyroscope
Varying the above parameters, simulation results were obtained for following
cases:
Translational Velocity of center of bar
Angle turned by the biped about the vertical axis

4.1. Casel: Effect of the mass of Gyroscope:

It is observed in Fig. 4 that the angle by which the biped turns is a function of the
mass (inertia) of biped, as shown analytically (refer equation 12). As the mass is
increased the corresponding values increase but not considerably. This shows
that the mass of the gyroscope does not affect the angle significantly and also
validates the parabolic nature of variation of angle with respect to time.

4.2. Case2: Effect of the angular velocity of Gyroscope:

Fig. 5 shows the variation of translational velocity of the centre of mass of the
bar with time, with changes in angular velocity of gyroscope. It is observed that
the points are more spaced out, and there is considerable effect of gyroscope
speed on translational velocity.
Fig. 6 shows the variation of vertical rotation of left leg with time, upon variation
in speed of gyroscope. The plots are linear in a large interval of time. This
interpretation will be helpful in the design of the biped in terms of the flywheel
speed.
 287

lime (SEG)

Fig.4 Variation of Vertical Rotation with time, for varying mass of Gyroscope

, , , * I

. * , I

,, ,, ., ,,
, . , , I

2400 0.01 o.m o.m 0.04 0.05 0.06 0.07 0.08 0.m 0.1
Time (sec)

Fig. 5 Variation of TranslationalVelocity with time, for different speeds of Gyroscope

Time (sec)

Fig.6 Variation of Vertical rotation about Left Leg with time, for different speeds of Gyroscope
288

5. Conclusion and Scope for future work

It is observed by means of simulation that gyroscopic effect can be used as
means of propulsion in bipedal locomotion. This paper does not attempt to
provide a means to simulate anthropomorphic gait, however aim of this paper is
to explore a phenomenon, and obtain an alternative means of bipedal
locomotion, using gyroscopic effect.
Future work would deal into analytical dynamic analysis of the system, and
experimentation on a prototype of the system. In addition to that, presently, no
concern has been kept in mind for the stability of the systems which is definitely
one of the main factors in bipedal locomotion.

References:
1. T. Mcgeer, IEEE International conference on Robotics and Automation, 3,
1640-1645, 1990.
2. B Barshan, and H.F. Durrant-Whyte, Proceedings of the IEEE/RSJ/GI
International Conference on Intelligent Robots and Systems 1867 - 1874
vo1.3, Sept 1994.
3. H.B Brown, Jr and Y.Xu, Proceedings of the IEEE International Conference
on Robotics and Automation, Vol. 4 , pp. 3658-3663, 1996.
4. J. Pratt, P. Dilworth, and G. Pratt. Virtual model control of a bipedal walking
robot. Proceedings of the 1997 International Conference on Robotics and
Automation, 1997.
5. M. Peck, Dynamics of a Gyroscopically Stable Hopping Robot, Proceedings
of the 2001 AAS Spaceflight Mechanics Conference, Santa Barbara, CA;
Feb. 200 1.
6. S .H Collins, M.Wisse, A. Ruina, “A Three-Dimensional Passive-Dynamic
Walking Robot with Two Legs and Knees” International Journal of Robotics
Research, Vol. 20, No. 2, Pages 607-615,2001.
7. MSC Software’s ADAMS View, 2005.
 A Self-Adjusting Universal Joint Controller for Standing and
Walking Legs

A . Schneider', B. Fischer, H. Cruse and J. Schmitz

Dept. of Biological Cybernetics, University of Bielefeld,
33501 Bielefeld, P.O.B o x 10 01 31, G e r m a n y
*E-mail: axel.schneider@uni-bielefeld.de
www.uni-bielefeld.de/biologie/Kybernetik/

Recent biological studies on stick insects suggest that the character of the joint
controllers (I-, P- or D-controller) depends on the compliance of the substrate
(soft, intermediate or inelastic) the insect is standing on. To model these results,
we propose a self-adjusting joint controller that changes its own setpoint in
dependance of the substrate stiffness. The substrate stiffness is determined
by means of a correlator circuit that compares injected movement commands
with the actual responses of the leg joint. The negative feedback controller
for standing is combined with an already published positive feedback loop for
walking to form a universal controller which can be used for the control of
legged robots.

Keywords: adaptive control, adaptive behavior, posture control, leg control

1. Introduction
Different biological studies provided evidence that in a walking stick in-
sect a positive feedback effect (assistance reflex) on the single joint level
is responsible for the generation of the stance movements and the coordi-
nation of all leg joints.6 It could be shown that legs equipped with local
positive velocity feedback joint controllers (LPVF controllers) are capable
of walking without a central control instance.8
In contrast, during standing each joint angle is controlled by a negative
feedback position controller (resistance reflex)'ig in order to maintain its
posture. However, a recent study by Cruse and co-workers on the control
of the knee joint of the standing stick insect changes the view on this topic
drastically. It suggests that the behavior of the negative feedback joint
controller depends on the compliance of the substrate the insect is standing
011.~Figure l(a) shows the setup of the underlying biological experiment.

289
290

joint controller

Fig. 1. (a) Design of the biological experiment. A stick insect is tethered t o a balsa
wood rod. The left middle leg stands on a small platform that is supported by a spring
steel beam the base of which can be moved away or towards the body on a perpendicular
path with respect t o the body axis (lower scale). At the same time the leg position is
monitored (upper scale). (b) Simplified model of the experiment.

A stick insect is tethered to a balsa wood mount. The left middle leg stands
on a small platform which is mounted on a spring steel beam. The base of
the beam can be moved sidewards either away or towards the insect body.
Since the spring constant ks of the spring steel is known, the force acting on
the tarsus of the leg in horizontal direction can be calculated by measuring
the tarsus position (pt,,,,,) and the position ( p s ) of the spring steel beam
(f = (Ps - Ptarsus) * ks).
On highly elastic substrate the negative feedback control circuit be-
haves like an I-controller. Externally enforced deviations from the original
position are compensated completely up to a spring steel stiffness of about
k , = 0.05 N/m. On a substrate with intermediate elasticity, up to a spring
steel stiffness of about; 1N/m, the animal compensates for the position only
partially which resembles the behavior of a P-controller. On inelastic sub-
strate the knee joint seems to be controlled by a D-controller. If the tarsus
is forced away passively from the original position, the controller tries to
compensate for this deviation. However, after a short period of time the
controller seems to “give up” and the new position is “accepted”. This
behavior is biologically sensible as it prevents the insect from exhaustive
torque generation for useless actions.
The present paper introduces a new self-regulating negative feedback
joint controller that generates the described behavior of the standing ani-
mal. Moreover, an architecture for the combination of both positive feed-
back for walking and negative feedback for standing is proposed.
 291

2. Simulation of the standing experiment

Figure l(b) displays the simplified setup of the experiment shown in
Fig. l(a). The joint angle -yj is measured (femoral chordotonal organ) and
serves as the only input of the negative feedback joint controller. It is as-
sumed that the negative feedback joint controller issues the motor command
T~ to the muscles (e.g. flexor tibiae) in order to achieve a certain angular
position of the joint. If an external force acts on the tarsus, there is a dif-
ference between -yj and -ym which is the angle of bending yb (-yb = yj - -ym)
due to the elasticity of the muscles and the tendons.
The angle of bending ^(b is also a measure for the amount of torque
I-, that acts on the joint (T, = -yb. k,, with k, = spring constant of the
elastic knee joint). The tarsus rests on top of the spring steel beam the base
of which is moved horizontally in the experiment. The horizontal distance
between the position of the spring steel beam and the position of the leg is
multiplied with the spring constant k, of the substrate in order to calculate
the force that acts upon the leg. The experiment was simulated in Simulink
6.1 and SimMechanics 2.2.1 (The Mathworks Inc., Natick, MA, USA) using
the kinematic data of an average stick insect as described by Ekeberg and
c o - ~ o r k e r s .The
~ joint controller which was applied in the simulation is
introduced in the following section.

3. A self-regulating negative feedback joint controller

In technical systems controllers are often designed as PID-controllers. Bio-
logical systems most often employ P- or PD-controllers.2 From a technical
point of view the control behavior of the knee joint controller in the stand-
ing stick insect is an apparent exception because it changes the controller
characteristic and the reference value of the controller according to the
elasticity of the substrate which is an environmental parameter. An ap-
parent design approach towards a controller that regulates the knee joint
as described above would be a PID-controller the parameters of which are
changed at runtime according to the sensed substrate e l a s t i ~ i t yHowever,
.~
such an approach would ignore an important aspect of the biological re-
sults. If the base of the spring steel beam is moved and the spring steel is
very inelastic, the leg is forced into the new position. The counter force of
the leg decreases after some time (D-controller) . But additionally, the joint
controller seems to accept the new position as reference value because after
a while a newly induced deviation from the latter position invokes the same
actions of the controller as that from the first position.
292

- -.I
I
I
I

I'-*-'-.-.

SimMechaniwmodel I
of leg on supporting
swna steel beam
.I
L J
I

Fig. 2. Combination of the joint controller for standing (bottom) and an LPVF con-
troller for walking (top, chain-dotted rectangle) as derived in Schneider et aL8

The controller introduced in this paper takes all the requirements men-
tioned so far into account. The block diagram is shown in Fig. 2. The core
of the overall control circuit is an I-controller (center of Fig. 2) which is
part of the negative feedback position control circuit that is indicated by
the area shaded in dark gray. The I-controller provides the motor command
ym and receives the difference between the actual joint angle yj and the
reference value y3,*@fas input.
The area shaded in light gray on the right side of Fig. 2 represents the
joint bending calculation, the resulting torque and the measurement of the
joint angle. The joint torque rr actuates the knee joint in the SimMechanics
simulation of the leg. At the same time the simulation provides the actual
joint angle yj. Taken together, the two areas shaded in dark and in light
gray generate a negative feedback position control behavior of the knee joint
that compensates completely for an externally enforced deviation from the
desired reference position. The basic function of the circuit therefore is that
 293

of an I-controller working in the joint angle domain.

The reference value that is passed to the I-controller is produced by
the reference integrator. The reference integrator acts as a memory for the
current reference value Yj,ref of the knee joint angle. Any input value of the
reference integrator that differs from zero shifts the reference value.
In order t o sense the elasticity of the substrate, a test signal Ytest is
generated (sinus function, f=70Hz, amp=O.l") and added to the motor
command output of the I-controller. This test signal leads to oscillations
in the joint driving torque T? and thus also to periodical fluctuations in
the angular position rj of the knee joint. However, the amplitude and the
phase of the superimposed oscillations in the knee joint angle depend on
the elasticity k, of the substrate. An infinitely stiff substrate would cancel
out the oscillations in the joint angle completely whereas a very soft ground
would allow the joint to follow the superimposed oscillations.
The correlator in Fig. 2 receives both the test signal -ytestand the angu-
lar response -yj of the knee joint. It is inspired by the Hassenstein-Reichardt
~ o r r e l a t o rThe
. ~ output of the correlator is calibrated to deliver values be-
tween 0 (ks = 00, inelastic substrate) and 1 (ks = 0, elastic substrate). In
principle, the measure for the correlation is obtained by multiplication of a
delayed version of the test signal Ytest (low-pass filter with the same time
constant as the leg joint) and the actual joint movement yj. The absolute
value of this product is filtered and represents the desired correlation value.
In order to bring the control circuit from the I-control to the P-control
mode, the upper of the two sigmoid functions (I/P sigmoid function) is used.
If the correlator output is 1, representing a soft substrate, the output of the
sigmoid function is zero. If the correlation decreases (stiffer substrate), the
output of the I/P sigmoid function increases. In this case, the product of the
angular position error ey (input of the I-controller) and the output of the
I/P sigmoid function is not zero anymore. The product is fed as input into
the reference integrator to move the reference value towards the actual joint
angle. In this mode, the I-controller and the reference integrator compete
to make the joint angle error ey zero: The former by moving the actual
joint and the latter by changing the reference value. As a result at a given
medium elasticity of the substrate, the joint partially follows an external
deflection. The behavior resembles that of a P-controller.
To shift the behavior further from the P-controller mode to the D-
controller mode, the lower of the two sigmoid functions (P/D sigmoid func-
tion) is used. If the correlation is small indicating a stiff substrate, the
product of the bending angle Y b and the output of the P/D sigmoid func-
294

spring const. of substr. spring const. of substr. spring const. of substr.

0.1 Nlrn 0.5 Nlm 10.0 Nlm

f=70Hz

ln

0 1 2 3 4 5 6 7 8 9 10 0 1 2 3 4 5 6 7 8 9 10 0 1 2 3 4 5 6 7 8 9 10
(4 (b) (c)

6-
5-
,F = 12.2 pN, F, = 25.4 pN

:lb/
4- f=70Hz
f=70Hz

01

0 1 2 3 4 5 6 7 8 9 1 0 0 1 2 3 4 5 6 7 8 9 1 0 0 1 2 3 4 5 6 7 8 9 1 0
time [s] time [s] time [s]
(d) (el (r)

Fig. 3. (a)-(c) Position of the Spring steel beam and leg position for external deflections
on three different substrates. (d)-(f) Leg forces for the same external deflections as in
(a)-(c).

tion increases. Since this product is also fed as input into the reference
integrator] it also shifts the reference value towards the actual joint angle.
This is done as long as a joint bending (torque) is present. As a result, the
controller “gives up” and does not produce any force after some time. This
completes the description of the adaptive joint controller for standing. The
top part of Fig. 2 is used for the control of the stance during walking (see
Sect. 5).

4. Results
The controller introduced in the last section was tested on a simulation of
an insect leg as described in Sect. 2 and Fig. l(b).
Figures 3(a)-(c) depict the leg position and the position of the base
of the spring steel beam during a deflection plotted over time for three
different substrate elasticities, exemplarily. Figure 3( a) displays a plot of
the position of the spring steel beam (gray) and the leg position (black) over
time for a soft substrate with a spring constant of 0.1 N/m. The deflection
 295

function for the beam position has an amplitude of 5mm. It can be seen
that the leg has the tendency to follow the beam position a t the beginning
of the deflection. However, the I-controller restores the original position
completely already during the course of the deflection. The I-controller
domain ranges from substrate elasticities of N/m to 0.2N/m. The
inset indicates the oscillations which were superimposed on the leg position
by the test signal TtYtest.The peak-to-peak amplitude of this oscillation is
34.3 pm.
Figure 3(b) shows the beam and leg position over time for a substrate
with a spring constant of 0.5N/m (intermediate elasticity) and the same
deflection function for the beam position as in Fig. 3(a). The joint controller
is in the P-control domain which leads to a passive movement of the leg of
about 1mm. The P-controller domain ranges from substrate elasticities of
0.2 N/m to 5 N/m. The oscillations superimposed on the leg position have a
smaller peak-to-peak amplitude (24.0 pm) for the given substrate elasticity
than for the softer substrate in Fig. 3(a).
Figure 3(c) presents the same plots as in Fig. 3(a) and (b) for a rather
inelastic substrate (lO.ON/m). The joint controller is in the D-control mode.
The leg follows the imposed deflection completely. The D-controller domain
ranges from 5 N/m to 100N/m. The superimposed oscillations are smaller
than in Fig. 3(b). They have a peak-to-peak amplitude of 2.5pm.

Figures 3(d), (e) and (f) show graphs of the leg forces during the simu-
lated experiments plotted over time for the I-, P- and D-mode of the con-
troller, respectively. The substrate elasticities are the same as in Fig. 3(a),
(b) and (c). Note that in Fig. 3(f), which represents a simulation run in
the D-domain, the controller generates a force at the onset of the deflection
with a force maximum a t about the end of the deflection ramp a t 6 s. How-
ever, this force decreases after a few seconds due to the effect of the P/D
sigmoid function. The insets in Fig. 3(d), (e) and (f) show the superimposed
force oscillations which are caused by the test signal. Note that opposed t o
the superimposed leg oscillations in Fig. 3(a), (b) and (c) the force oscil-
lation increase with rising substrate stiffness. This is due to the fact that
the test signal generates oscillating motor commands that produce higher
forces when the leg’s freedom to move is decreased by a stiffer substrate.
Figure 4 depicts the leg position plotted over the leg force. For a sub-
strate elasticity of 0.1 N/m the leg force increase up to 1mN without chang-
ing the leg position (I-control). In the P-control mode (substrate elasticities
of 0.2 N/m t o 5 N/m) the simulation shows an almost linear relationship
N
296

0 1 2 3 4 5 6 7
leg force [mN]

Fig. 4. Leg position plotted over leg force. Values are taken from the end of simulation
cycles when the system has reached a steady-state.

between force and position which reflects the spring like behavior of a P-
controller. The maximum force is achieved for a substrate elasticity of about
1N/m. In the D-domain (substrate elasticities of N 5 N/m to 100 N/m and
more) the static forces decrease again until the leg follows the deflection of
the base of the beam completely.

5 . Combined controller
In previous publications, the control principle of Local Positive Velocity
Feedback (LPVF) and its derivatives have already been i n t r o d u ~ e d . The
~!~
LPVF controller is used to generate and maintain movements in closed
kinematic chains without a central controller. This is the case for a leg
in stance during walking. This section presents a combination of both the
LPVF controller for walking and the self-regulating negative feedback joint
controller for standing. The block diagram of the combined controller is
shown in Fig. 2.
The component that is responsible for the generation of the stance move-
ment during walking (LPVF) can be found in the upper part of the figure
(enclosed in the chain-dotted rectangle). The component of the controller
that is active during standing is shown in the lower part (cf. Sect. 3). The
combined controller uses the reference integrator and the I-controller (cen-
 297

ter part of Fig. 2) for both modes, i.e. standing and the generation of stance
movements. The toggle switch S2 selects between the positive velocity feed-
back branch (top) and the negative position feedback branch (bottom). The
two behaviors depend on the activation state of the animal and might be
chosen by a central instance.
If the LPVF branch is selected t o perform a stance movement, the ac-
tual joint velocity Ayj is calculated by the differentiator (top middle) from
a series of joint angle inputs yj.The angular velocity Ayj is fed into the
input of the reference integrator via the summation point (top left) and
the switch S2 forming the actual positive feedback loop. Whenever a joint
movement is initiated, the positive feedback loop via the differentiator main-
tains and continues this movement actively. However, if not coordinated,
the resulting joint movement during a stance phase leads t o tensions in all
joints of the legs on ground which form so called closed kinematic chains.
In classical, central controllers these tensions are avoided by issuing only
those motor commands that lead t o coordinated movements of all joints.
In the LPVF approach the minimization of joint tension (and hence coor-
dination) is achieved by an active relaxation strategy on the single joint
level. The power controller (top left of Fig. 2) calculates the actual joint
power which is the product of the angular joint velocity and the joint torque
(Py = Ayj . ry = Ayj . yb Icy). The switch S1 is toggled by the power con-
+

troller depending on the degree of positive or negative mechanical power

generated in the joint. If the mechanical joint power is negative, the joint
torque and its corresponding joint movement point into different directions.
The joint is in coasting mode which indicates unnecessary tension. There-
fore, the tension is relaxed by closing the switch S1 which leads to a shift
of the reference integrator input by the actual amount of bending (bending
angle yb). This can be regarded as an active relaxation of the joint (active
compliance). However, if the mechanical joint power is positive, the joint
is in traction mode which leads to a meaningful movement of the entire
walking agent. In this case, the switch S1 is left open and the ongoing joint
movement is not changed. Instead of using a discrete switch S1,the active
relaxation can also be achieved by a continuous mechanism that fades in
the bending signal according to the amount of (negative) mechanical power
that is generated in the joint.

6 . Discussion
The combination of the self-regulating negative feedback joint controller
and the LPVF controller results in a local controller that regulates a joint
298

in a leg during standing and during a stance movement. The negative feed-
back joint controller is active during standing and is responsible for the
maintenance of a certain angular position in order to keep a desired pos-
ture against external disturbances but not at any (torque-) costs.3 Instead,
the leg joint maintains a joint angle when standing on soft substrate and
ceases if the leg is pulled away on a rigid substrate. The former might be
the case if a leg is placed on a moving leaf, the latter if the leg is placed on a
rigid branch moved by the wind. The new self-regulating negative feedback
joint controller can explain the different behaviors.
In combination with the LPVF controller for walking,s a universal, self-
adjusting joint controller for standing and walking can be built. Besides
the agent’s internal impulse to change between standing and walking, tog-
gling between the two controller types might also depend on the correlation
measure or on the angular acceleration for example t o avoid slipping.

References
1 . U. Biksler. Neural Basis of Elementary Behavior in Stick Insects, volume 10
of Studies of Brain Function. Springer, Berlin, London, New York, 1983.
2. H. Cruse. Neural Networks as Cybernetic Systems. Thieme, Stuttgartl New
York, 1996.
3. H. Cruse, S. Kuhn, S. Park, and J. Schmitz. Adaptive control for insect leg po-
sition: Controller properties depend on substrate compliance. J. Comp. Phys-
i ~ l . 190:983-991,
, 2004.
4. 0. Ekeberg, M. Blumel, and A. Buschges. Dynamic simulation of insect walk-
ing. Arthropod Struct. Devel., 33(3):287-300, 2004.
5. B. Hassenstein and W. Reichardt. Systemtheoretische Analyse der Zeit-,
Reihenfolgen- und Vorzeichenauswertung bei der Bewegungsperzeption des
Riisselkafers Clorophanus. Z. Naturforsch., B11:513-524, 1956.
6. J. Schmitz, C. Bartling, D.E. Brunn, H. Cruse, J. Dean, T. Kindermann,
M. Schumm, and H. Wagner. Adaptive properties of “hard-wired” neuronal
systems. Verh. Dtsch. 2001.Ges., 88(2):165-179, 1995.
7. A. Schneider, H . Cruse, and J. Schmitz. A biologically inspired active compli-
ant joint using Local Positive Velocity Feedback (LPVF). IEEE Tkunsactions
on Systems, Man, and Cybernetics - Part B: Cybernetics, 35(6):1120-1130,
2005.
8. A. Schneider, H. Cruse, and J. Schmitz. Decentralized control of elastic limbs
in closed kinematic chains. The International Journal of Robotics Research,
Special Issue on CLAWAR 200.4, 25(9):913-930, 2006.
9. G. Wendler. Laufen und Stehen der Stabheuschrecke Carausius morosus: Sin-
nesborstenfelder in den Beingelenken als Glieder von Regelkreisen. 2. Vergl.
Physiol., 48:198-250, 1964.
 A STEP TOWARDS PNEUMATICALLY ACTUATED BIPED
LOCOMOTION :A BIO INSPIRED PLATFORM FOR
STIFFNESS CONTROL

G.MUSCAT0, G. SPAMPINATO
Dipartimento di Ingegneria Elettrica Elettronica e dei Sistemi (D.I.E.E.S.)
Universitir degli studi di Catania, Viale Andrea Doria 6, 95125 Catania, ITALY
smirscato@,diee,s.unict.it , .sspampi@,iees.unict.it

This paper presents some control methodologies in the field of biped locomotion. Based
on both two and three dimensional simulations, the control strategy has becn
implemented upon a ten degrees of freedom biped robot pneumatically actuated. The
locomotion strategy has been developed using a simple but efficient force and position
control algorithm, based on a different interpretation of the Virtual Model Control
approach. Some energy considerations has been also carried out in order to better tune the
gait parameters.

1. Introduction
The study of the walking strategy in biped robots gives more accurate insight
into the strategy actuated by human beings in keeping balance during gait. Up to
now only a few humanoid robots have been completely realize6 like the famous
Honda humanoid robots [ l ] or the Sony Robot Qrio [2]. However, in order to
investigate upon the elastic properties of human legs, some other pneumatic
biped robots have also been realized, such as the Bipedal Robot LUCY [3], or
the D.I.E.E.S. biped robot shown in Fig.1, which has been entirely realized in
the laboratory of the University of Catania. In the present work pneumatic
actuators have been adopted as an alternative solution to the most widely used
electric actuators. Due to their low weight-force ratio and their high elasticity,
pneumatic actuators allow the controller to adapt the stiffness of the leg during
the stance phase. The D.I.E.E.S. biped robot is a 10 Dof anthropomorphic biped
pneumatically actuated. The robot is 110 cm tall, and weights 24 kg (including 8
Kg payload). The design of the shape and the dimensions of the single parts are
human inspired, so the dimensions of the links and the articulation movements,
reproduce the corresponding biological ones [4]. Two pneumatic cylinders
actuate the ankle and the hip articulations, while the knee articulation is actuated
by only one. The ankle and the hip articulations are made up by a cardan joint

299
300

that provides both roll and pitch degrees of freedom. Moreover, the two
degrees of freedom in each articulation are strictly coupled with the two
pneumatic actuators motions. In other words, a single movement of a single
degree of freedom involves the extensions of both the pneumatic cylinders and
vice versa. Different kind of algorithms have been implemented in order to
solve the inverse kinematical problem. Some static maps, using the neural
network approach, and dynamic algorithms have been proposed and widely
compared in [ 5 ] .

Figure 1. The D E E s Biped Robot

2. L o ~ o ~ o t i ostrategy
n and §i~ulations
The locomotion strategy presented makes use of both force and position control
approaches in order to actuate the stance leg and the swing leg to support the
robot weight and to place the swing foot in the ahead position planned on-line
by the trajectory generation algorithm. In order to generate the stance leg joint
torques either in the Single Support Phase (SSP) and in the Double Support
Phase (DSP), a set of virtual forces have been applied to the hip according to a
different interpretation of the Virtual Model Control concept [6],[7]. In
particular, a virtual spring damper couple have been applied between the hip
joint and the stance foot in contact with the ground, applied to the desired COP
position, as shown in Fig.2. In this way, each leg is considered as a serial
manipulator, and the forces exerted by the foot, are directed in line with the
robot center of mass (approximately located in the robot pelvis), preventing the
body to rock and minimizing the cost of transport [S]. This approach represents
 301

a different inte~retation of the VMC concept described in [6]. The

implementation is quite different because no underactuated joint is required to
model the foot-ground contact and no foot ground inclination measurements are
needed.

Figure 2. The two dimensional walking pattern based on the virtual spring damper components
applied on each leg.

Besides, some analytical conditions have to be formalized in order to simplify

the joints torques generation conesponding to the virtual forces exerted. In other
words, taking into account the complete 3x3 leg jacobian matrix J L ~some ~,
conditions have been demonstrated in order to guarantee the force exerted
connecting the hip joint to the desired COP position as shown in Fig.3.
tY

Figure 3. Two dimensional external force exerted by the stance leg to the desired COP point.

The necessary and sufficient condition (CNS) shown by (1) has been
demonstrated, in which the terms JP represents the 2x2 jacobian sub-matsix
containing the first and the x-th columns of JLer

Moreover, in order to simplify the application of the condition (I), the force
intensity lFExt/has been directly considered. So, the torques needed to guarantee
302

the alignment condition and satisfying the CNS (1) can be calculated through
the relations (2).

.------.
det J:

The paper aims also to describe a theoretical method to find out some virtual
components parameters according to the energy constraints and natural walking
pattern. The total mechanical energy of the system can be computed as the sum
of the kinetic and the potential energy terms, where the last one includes both
the gravitational and the elastic energy stored into the virtual springs. For
simplicity it is assumed to investigate the energy expenditure of equivalent
virtual components acting in the vertical direction in order to support the robot
weight during the stance phase. So, the mechanical energy stored in the system
can be expressed as in (3).

The main constraint taken into account is the maximum energy that can be
supplied to the system Emax. The desired COG height from the ground and the
elastic spring constant are also key parameters to be considered. So, three
parameters related to each other by the expression (3) regulates the robot
dynamics. Once fixed the maximum energy allowed Emax, and the
uncompressed spring length ho, the system potential energy can be calculated
and represented with respect to the hip y-coordinate, as shown in Fig.4.
<
Energy (Joule1
FnslnY ln,,lsl EWQYIJouIel
EnmY IJouIeI

s 03 02 01 0. 0, 0% 07 0% OD I
y n v :m~ Y HIP :m1

Figure 4.Gravitational and elastic potential energy with respect to the hip y-coordinate.
 303

In the left part of the Fig.4, both elastic and gravitational potential terms are
represented together with their sum U(YHjp).A zoom is reported in the right side.
The motion of the robot hip has to remain within the segment AB. Indeed,
outside this segment, the motion is impossible because of the potential energy
exceeding the maximum energy allowed. In order to chose a proper spring
elastic constant, the constraint (4) has to be satisfied.

min&,,) = u,, 5 E,, (4)

Deriving the U(YH*) expression shown by (3) with respect to Yhjp , the minimal
potential energy point indicated by Urnin
can be found.

u,, =-- (mg )* + mgh (5)

2K
Finally, substituting the equation ( 5 ) into (4) and solving for K the desired
virtual spring constraint has been found.

The energy supplied to the system E M x has been fixed to 320 J, and the
uncompressed spring length ho to 0.8 m. The K parameter has also been chosen
in order to maintain the energy requirements and to avoid too wide oscillations
around Yrnj,,.In other words, the segment AB length corresponding to decreasing
values of K increases, and the steady state hip height Ymindecreases as shown in
the left part of the Fig.5.

Figure 5. Minimal potential energy hip position related to the chosen value of the virtual spring
constant, and related oscillations during the simulation.

Some simulations have been performed according to the virtual component

criteria, where the minimal potential energy hip position has been chosen to be
304

about 70 cm high from the ground. In the meanwhile, the swing leg, when not
directly involved in balancing the robot during the Double Support Phase, is
controlled in position, in order to perform parabolic trajectories, like described
more in depth in [7]. Some simulation results are reported in Fig. 6.

-9$ B m ? w * m = . 3 ? a Q . m
r2imr

Figure 6 . Two dimensional simulation results, ZMP, left and right feet positions during the gait.

After some rearrangements of the relations (1) to the three dimensional case,
some equivalent equations have been found like shown by the condition (7). In
other words, in order to simplify the generation of the virtual forces under the
alignment condition constraint, the Jacobian matrix J30.f related to each leg is
calculated during the gait, according to the robot posture and the desired COP
position.

I zK=-
det JCIz3
‘A-roN det Jc124

(7)
M , = M y= M , =o
-
‘H-roll - ‘H-pttch =’

Again the magnitude of the exerted force lFExll has been directly considered, and
the torques needed to guarantee an equivalent CNS for the three dimensional
case are shown by the relations (8).
The JR *y terms represent the 2x2 sub-matrix containing the x-th and the y-th
rows of the JLeg matrix, while the terms J$2sx represents the 3x3 jacobian sub-
matrix containing the first, the second and the x-th columns of JLer Posing
TH-, . o ~ / = ~ ~ g i l c h = O and the other joints torques according to (8), the alignment
condition shown in Fig.7 for the virtual stance forces is automatically
guaranteed.
 305

Also for the three dimensional case, some simulation have been carried out in
order to validate the locomotion strategy based on the virtual components. Like
for the two dimensional case, the swing foot is also driven directly into the three
dimensional space through parabolic trajectories. A 60 cm step length
locomotion pattern is reported in Fig.7.

Figure 7. Three dimensional simulationsbased on the wrtual force generation for the stance leg and
parabolic trajectones for the swing leg.
306

3. Cantrol architecture
As discussed in the previous section, the stance leg is actuated in order to exert
the needed force to support the robot directly to the COP point, whereas a three
dimensional trajectory generation has been designed for the swing leg in order
to drive the swing foot directly into the operative space. So, the control
architecture makes use of a three level feedback loop controller shown in Fig.8

W C QeireraUon

Figure 8. The three level control architecture.

The inner control loop operates on the pneumatic actuators directly performing
both the position and the pressure loop controller. The second level controller
works on the articulation kinematics and kinetics, in order to provide the
actuators lengths and forces corresponding to the joints positions and torques
generated by the high level controller. The feedback information are the joint
working angles, generated from the kinematic inversion algorithm. The third
level controller represents the highest level controller, where the trajectory
generation and the stance leg forces co~putationare performed, based on the
sensors information.
In order to measure the COP position inside the foot contact surface, each leg
have been equipped with an inductive pressure sensor placed under the feet
soles. The sensor structure is made up of two main parts, with an elastic
dielectric material between them [4], as shown in Fig.9. Finally, the 2.0 €3 active
version of the CAN protocol has been implemented at 1 h4b in order to connect
the sensors and the actuators control boards to the main controller.
rnwum

Figure 9 Thc foot pressurc scnsor structurc

 307

4. Experimental results and conclusions

Some experimental walking tests have been performed on the real robot in order
to validate the walking strategy. So, a 30 cm walking pattern locomotion test is
shown in Fig.10. Apart from the virtual component parameters estimated
through energetic cost considerations, the tuning process for other walking
parameters like the lateral displacement or the step length are still under
investigation in order to guarantee the dynamic stability conditions for robust
and autonomous walking patterns.

Figure 10. Two 30 cm step length locomotion tests performed on the real robot.
308

References
1. J. Chestnutt, M. Lau, G. Cheung, J. Kuffner, J. Hodgins, and T. Kanade,
“Footstep planning for the honda asimo humanoid,” in Proceedings of the
2005 IEEE International Conference on Robotics and Automation
Barcelona, Spain, 2005.
2. K.Nagasaka, Y .Kuroki, S.Suzuki, Y .Itoh,, J.Yamaguchi. “Integrated
Motion Control for Walking, Jumping and Running on a Small Bipedal
Entertainment Robot”. Proceedings of the IEEE International Conference
on Robotics and Automation, New Orleans, April, 2004.
3. B. Vanderborght, B. Verrelst, M. Van Damme, R. Van Ham, P. Beyl, D.
Lefeber, “Locomotion Control Architecture for the Pneumatic Biped Lucy
consisting of a Trajectory Generator and Joint Trajectory Tracking
Controller,” in Proceedings of the 2006 6th IEEE-RAS International
Conference on Humanoid Robots, Genova, Italy, December 4-6,2006.
4. G.Muscato, G. Spampinato, “A Pneumatic Human inspired robotic Leg:
Control Architecture and Kinematical Overview”, International Journal of
Humanoid Robotics, Volume 3, Number 1, pp.49-66, March 2006.
5. G. Muscato, G. Spampinato, “Kinematic Model and Control Architecture
for a Human Inspired Five DOF Robotic Leg”, International Journal of
Mechatronics, Volume 17 Number 1, pp. 45-63, February 2007.
6. J.Pratt, C-M Chew, A.Torres, P.Dilworth, G.Pratt. “Virtual Model Control:
an intuitive Approch for Bipedal Locomotion”. The International Journal of
Robotics Research. Vol. 20, No, pp 129-143,2, February, 2001.
7. G. Muscato, G. Spampinato, M. Costa, “Virtual Forces Based Locomotion
Strategy and Energy Balance Analysis” Proceedings of the 8th International
IFAC Symposium on Robot Control SYROCO 2006 Santa Cristina
Convent, University of Bologna (Italy) September 6 - 8,2006.
8. R. M. Alexander, Principles of Animal Locomotion. Princeton University
Press, 2003.
AUTONOMOUS BIPEDAL WALKING GAIT ADJUSTMENT
UNDER PERTURBATIONS

L. YANG C. M. CHEW and A. N. P O 0

Department of Mechanical Engineering, National University of Singapore,
Singapore, 119260
E-mail: yanglin, mpeccm, mpepooanQnus. edu.sg
http ://gupp y .mpe .nus. edu.sg/ mpeccm/chew cm.shtml

This work focuses on the walking stride-frequency autonomous adjustment in

response to the environment perturbations. Reinforcement learning is assigned
t o supervise the stride-frequency. A simple momentum estimation further as-
sisted the adjustment. In the learning agent, a sorted action-choose table in-
structed the learning to find out the proper action in a straightforward way.
Incorporating the real-time steplength adjustment mode, the presented gait
adjustment is able to achieve different pace walking adaptive to the environ-
ment. Dynamic simulation result shows the supervision is effective.

Keywords: Q learning CMAC network real-time walking stride-frequency walk-

ing step-length

1. Introduction
The ultimate tangible goal of the research presented here is to achieve
stable bipedal walking adaptive to the environment using recursive search
and computations with accessible robot's dynamics.
Bipedal locomotion control is a highly nonlinear and multi-aspect con-
trol issue. Although some research work has been shown effective for giving
a stable walking, walking adaptively to the environment change is still a
challenge for further explorations. Among the developed methods, Central
Pattern Generator (CPG) concept inspired methods are expected to be
flexible with environment. Current popular CPG models such as neural os-
cillators' and Van der Pol equations2 use nonlinear coupled equations to
elaborate the leg harmonious patterns. However, so far it still has not been
clearly depicted about its feedback pathways and the relationship between
the oscillator models and robot dynamics.
The presented work investigated the issue about bipedal walking gait

309
310

real-time adjustment under perturbations. It is based on a newly proposed

motion generator Genetic Algorithm Optimized Fourier Series Formulation
(GAOFSF)3,4 which partitions a motion control into a low-level motion
generation, and a high-level motion supervision. The low-level part34 gen-
erated optimized walking solutions and contained parameters for online
a d j ~ s t m e n t For
. ~ the high-level motion surveillance, a reinforcement learn-
ing (RL) algorithm is used as the core agent to supervise the walking pace
to be adaptive to the enviornment. In this paper, we focus on the high-level
supervision for walking under perturbations. The target robot is modelled
as the same as the one modelled in a previous work r e p ~ r t . ~

2. Stride-frequency online adjustment

This section discusses the automatic stride-frequency adjustment using the
Truncated Fourier Series (TFS) model. (1) and (2) gives the general pattern
generator equation for hip and knee joint respectively.
&+ = Cni=,R.Aisiniwh(t - t h + ) + ch
(1)

where Oh+ and oh- are positive and negative hip joint angles, respectively.
Ai, Bi and Ci are constant parameters determined by the Genetic Algo-
rithm, R is a scaling constant, and t h + , th-and tk are all time-shift values
(refer to our previous work3). ch is the defined hip offset value which indi-
cates the place where two hip joint trajectories intersect. ck indicates knee
joint's lock phase angle value.
Stride-frequency is defined as the fundamental frequency wh in the TFS
model. Since walking speed has been optimized even using the GAOFSF,
this stride-frequency can represent the average walking speed. The main use
of the stride-frequency adjustment is to maintain the kinetic energy under
perturbations. The accessible range of the adjustment can be referred to
our previous paper.5

2.1. Modules included in the Reinforcement Learning ( R L )

Reinforcement Learning (RL)is a class of learning problem in which an
agent learns to achieve a goal through trial-and-error interactions with the
environment. The learning agent learns only from reward information and
 311

it is not told how to achieve the task. From failure experience, the learning
agent reinforces its knowledge so that success can be attained in future
trials. The environment is modelled to be a fully observable Markov decision
process (MDP) that has finite state and action sets. Q-learning method
recursively estimates a scalar function called optimal Q-factors (Q*(i, u))
from experience obtained at every stage, where i and u denote the state
and corresponding action, respectively. The experience is in the form of
immediaterewardsequence,r(it,ut,it+l)(t = 0 , 1 , 2 . . .). ( Q * ( i , u ) )givesthe
expected return when the agent takes the action u in the state i and adopts
an optimal policy 7r* thereafter. Based on (Q*(Z,u)),an optimal policy
7r* can easily be derived by simply taking any action u that maximizes
( Q * ( i ,u)) over the action set V ( i ) .(3) gives the single-step sample update
equation for Q ( i ,u ) : ~
Qt+i+Qt(it,ut) + a t ( &Ut)[r(it,ut,it+i)
,

+ymaZ,EU(it+l)Qt(it+i,u)- Q t ( i t , ut)] (3)

where the subscript indicates the stage number; r(it,ut, & + I ) denotes the
immediate reward received due t o the action ut taken which causes the
transition from state it to it+l;Q E [O, 1)denotes the step-size parameter for
the update; y E [0, 1) denotes the discount rate. (1) updates &(it,ut)based
on the immediate reward r ( i t ,ut, and the maximum value Q(it+l,u)
of over all u E U(ti+l).
The convergence theorem for the Q-learning algorithm that uses lookup
table representation for the &-factors is: Theorem (Convergence of Q-
1earning):For a stationary Markov decision process with finite action and
state spaces, and bounded rewards r(it,ut, & + I ) If the update equation (1)
is used and at E [0, 1) satisfies the following criteria: Czl at(i,u) = 00
and C ~ l [ a t ( i , u < ) ]00,V(i,u);
2 then Qt(i,u)+Q*(i,u) as t -+ 00 with
probability 1, V ( i , u).
In this paper, Cerebellar Model Articulation Controller (CMAC)7 is
used as a multi-variable function approximator for the Q-factors in the
Q-learning algorithm. CMAC has the advantage of having not only local
generalization, but also being low in computation. The Q-learning algorithm
using CMAC as the Q-factor function approximator is summarized as:
Initialize weights of CMAC
Repeat (for each trial):
Initialize i
Repeat (for each step in the trial)
Select action u under state i using policy (say E ) based on Q.
312

Take action u,

-+
Detect new state i‘ and reward r
i f i’ not a failure state
6 r yrnazdQ(i’, u’) - Q(i,u )
Update weights of CMAC based on b
i c a‘

6 -
else ( r = r f )
r f - Q(;,U )
Update weights of CMAC based on S
Until failure encountered or target achieved
Until target achieved or number of trials exceed a preset limit
State variables: The input state representing and describing the mo-
tion dynamics for the learning agent is composed of (1) trunk’s CG error;
(2) external force f ; (3) current stride-frequency w t ; the output state is
the normalized change of the stride-frequency Awh with respect to different
step-length scales. Such state variables contain the current average walking
velocity from the current stride-frequency for a fixed step-length walking
(supposing the change of stride-frequency is proportional to its step-length
level), motion acceleration which depicts the dynamic trend through ex-
ternal force and the actual stance posture obtained from the trunk’s CG
error.
Reward functions: reward functions give rewards to motions which
satisfied the motion objectives. However, it also punishes a failed action.
The reward function is shown as (2):
Ic . (C - IAwhI) for actions succeeded
reward =
{ -Ic,lEl - lctlTql for actions failed
(4)
Ic,lc,,lct are the scales for the performances. C is a positive constant which
stands for a fixed reward; E and Tq are the error of trunk’s CG and the
excessive ankle joint torque. For actions succeeded, there is still a differ-
ence considered among different actions. If the change of stride-frequency is
small, it means the updated frequency is not deviated much from the orig-
inal frequency, the action will be given more rewards. This is to encourage
the walking tolerance provided the selected action is successful.
Learning tasks: In the RL supervision, the more experiences the learn-
ing agent has gone through, the more reliability the controller can ensure.
The advantage is as long as the learning agent is established, it can always
keep learning different tasks assigned and increase the experience unbound-
edly. However, it is not easy to learn enough experiences in a short time.
To reduce the chance that the robot might encounter situations which have
 313

not been trained yet, 6 representative tasks were assigned to the RL.In the
paper, flat-terrain walking3 is taken as an example.
Firstly, the stride-frequency W h and external force f are identified into
groups w s 1 w m , W b and f s l f m , f b . The subscript s , m , b stands for small,
medium, big, respectively. With the learning carried on, the following
groups can be further separated into subgroups.
W, : wh E [2.9,4.2] c)W t = 3.5 Wm : wh E [4.3,5.6] t--)W t = 4.9

f m :f E [30, 50) - f t = 40 f b : f E

The representative 6 tasks are described as:

[50,70) +--) f t = 60

1) wo = w, n f = random(f,) with the objective: continuous walking at

least for 4.0 seconds with the frequency increased to W b .
2) wo = ws n f = random(f,) with the objective: continuous walking
at least for 3.0 seconds with the frequency increased to w b .
3) wo = w, n f = m n d o m ( f b ) with the objective: continuous walking at
least for 1.2 seconds with the frequency increased t o w b .
4) wo = w b n f = -random(f,) with the objective: continuous walking
at least for 1.5 seconds with the frequency decreased to w,.
5) wo = W b f l f = -random(f,) with the objective: continuous walking
at least for 1.0 seconds with the frequency decreased to w,.
6) wo = W b f l f = -?-andom(fb) with the objective: continuous walking
at least for 0.7 seconds with the frequency decreased to w,.
Common constraints are: the absolute value of trunk’s CG error <
0.05m (trunk CG height is 0.925m); the sum of all joint torques in one dis-
crete time step (0.01s) is less than 1000ON.m (one discrete-step has 100-200
time refreshes); and the stance foot has no rotation bigger than loo. There-
fore, with the torque constraint getting strict, the energy consumption will
be expected reduced; with the extension of the walking time, the tolerance
will be improved.
Action-choose Table: To reduce the number of iterations and choose
the correct action easier, a logic instruction is established shown in Table
1.
Momentum estimation: This is to enhance the confidence for the
stride-frequency online adjustment in case the current situation has not
been experienced in the CMAC network. Therefore, it is only used when
the current state has not been trained by the CMAC network yet. The
stability issue and the allowed sudden change of the stride-frequency have
314

been discussed in the previous work.5 Therefore, as long as the estimated

change is within the allowed range, the dynamic stability of this estima-
tion is promised. According to the sensed or calculated external force and
the stride-frequency update time-step, the input momentum given by the
perturbation can be estimated by (3)to (5).

M = f t . At = C:=,mi. Avi (5)

Av4 = nu5 = n u s = nu7 = n u (7)

M is the input momentum, 11, 12 and L are the link-length with effect of
upper, lower and the whole leg respectively. i is the link number starting
from the stance foot t o the swing foot. The change of stride-frequency is
calculated using (6):

S is the step-length and Av can be calculated through (3) to (5). (6) actually
updates the walking t o be out of the effect of the external force if the adjust
is within the stable range. For the situation that the perturbation is opposite
t o the walking direction, stride-frequency should actually be kept t o sustain
the walking or call the backward motion. Nevertheless, the RL was applied
to both situations even if the latter is not so important.
 315

3. Step-length online update

The presented step-length adjustment is considered externally, not included
in the learning control. It is to adjust the position error. Step-length up-
dates itself through the scaling parameter R in the equation (1) and (2)
whenever the investigated phase angle has an error t o the phase defined in
the transition m ~ d u l e Actually
.~ changing the step-length is to maintain
the walking pattern according to the stance posture while tuning the stride-
frequency is to maintain the kinetic energy. Therefore, with the step-length
and stride-frequency online adjustment, robot dynamics becomes more con-
trollable under perturbation or transitions. The basic rule for adjusting the
step-length follows the linear i n t e r p ~ l a t i o n .The
~ further step-length ad-
justment mode will be presented in the future work.

4. Experiments on walking under perturbations

This part presented the learning result and dynamic simulation result of
the supervisory control applied to the bipedal walking under perturbations.
For the recovery of a pattern, as soon as there is no external force sensed,
the robot will be assigned the recovery module which follows the linear
approaching.

4.1. Reinforcement Learning Results

Fig. 1 and Fig. 2 show the Q-learning performances for the 6 tasks men-
tioned in Section 2.1. Since the basic walking pattern has already been
regulated, the RL learning task is relatively simple and not time consum-
ing yet.

4.2. Dynamic Simulation Experiment

The aforementioned algorithm has been implemented dynamically on a
walking experiment described as: external forces 20N, 40N, 60N were ap-
plied from t = 2s to t = 5s, t = 6s to t = 6.8s and t = 9s to t = 9.25s re-
spectively; external forces -2ON,-40N,-60N were applied from t = 11.2s
to t = 11.6s, t = 13.7s to t = 14.1s and t = 16s to t = 16.3s respectively.
With the presented RL agent and the momentum estimation. The robot
is able to overcome all batches of perturbation and behaves similar t o hu-
man beings. Fig. 3 shows robot walking was disrupted without any mo-
tion adjustment but Fig. 4 and Fig. 5 show the successful walking and the
real-time stride-frequency and step-length adjustment information under
all perturbations.
316

Learningtask 1

1
- learningtask 3

e 3
!
!
I

I
'0 10 20 30 40 50 60 70
Iteration

Fig. 1. Learning performance of task 1 to 3.

2
Learningtask 4 I

I
0' 5 10 15 20 25 30 35
Iteration

Fig. 2. Learning performance of task 4 to 6.

5 . Conclusion and future works

The designed learning agent, momentum estimation and step-length update
rule have been effectively applied to the motion supervision for biped's
stride-frequency and step-length adjustment. As motion pattern has been
regulated and optimized at low-level first using GAOFSF3,4 it should not
take a long time to train the RL network. Concluded from the conducted
 317

I
0 0.5 1 1.5 2 2.5 3 3.5

0 0.5 1 1.5 2 2.5 3 3.5

Trunk pmition (m)

Fig. 3. Stick diagram of walking without stride-kequency or step-length adjustment

(walking was disrupted finally).

Distance (m)

Fig. 4. Stick diagram of walking with online stride-frequency and step-length adjust-
ment (under lst, 2nd and 3rd batch of perturbations).

walking experiment, this supervisory control worked well and instructed the
robot t o get over of all batches of perturbations. Further discussions about
the analysis of the performance regarding less joint torque, trunk position
error and better robustness of this supervisory controller will be presented
in the following report.
318

E O5

0
I I
6 7 8 9

2
5
- -40
-60
6 6.5 7.5 8 8.5

7 75 85
Distance (m)

Fig. 5. Stick diagram of walking with online stride-frequency adjustment(under 4th and
5th perturbations).

References
1. Y. Fukuoka H.Kimura and A.H. Cohen Adaptive Dynamic Walking of a
Quadruped Robot on Irregular Terrain Based on Biological Concepts The
International Journal of Robotics Research(2003) Vol 22 No. 3-4 pp. 187-
202.
2. M.S. Dutra A.C. de P. Filho and V.F. Romano Modeling of a bipedal locomo-
tion using coupled nonlinear oscillators of Van der Pol Biological Cybernetics
(2004) Vol 88(4), pp 286-292 .
3. L. Yang C. M. Chew A. N. Po0 and T. Zielinska Adjustable Bipedal Gait
Generation Using Genetic Algorithm Optimized Fourier Series Formulation
IEEE/RSJ Int. Conf. on Intelligent Robots and Systems(2006) pp 4435-4440.
4. L. Yang C. M. Chew T . Zielinska and A. N. Po0 A Uniform Biped Gait
Generator with Off-line Optimization and On-line Adjustable Parameters
Robotica (2007) accepted.
5 . L. Yang C. M. Chew A. N. Po0 and T. Zielinska Adjustable Bipedal Gait
Generation using GA Optimized Fourier Series Formulation; Real-time Gait
Adjustment Proceedings of the Int. Conf. on the Autonomous Robots and
Agents (2006) New Zealand, pp 111-116.
6. John N. Tsitsiklis Asynchronous stochastic approximation and q-learning
Machine Learning (1994) Vol 16 (3) pp 185-202.
7 . Thomas W. Miller and Filson H. Glanz The university of new Hampshire im-
plementation of the cerebellar model arithmetic computer-cmac. Unpulished
reference guide to a CMAC program (1994).
 DESIGN AND PROBLEMS
OF A NEW LEG-WHEEL WALKING ROBOT

CRISTINA TAVOLIERI'
ERIKA OTTAVIANO
MARC0 CECCARELLI
L A M : Laboratory of Robotics and Mechatronics, University of Cassino
Via di Biasio 43, Cassino. 03043, Italy

In this paper an analysis is presented for new leg mechanisms of a 1-DOF (Degree-Of-
Freedom) leg-wheel walking machine. A preliminary prototype of a low-cost robot,
which is capable of a straight walking with only one actuator, has been designed and built
at LARM: Laboratory of Robotics and Mechatronics in Cassino. Simulation and
experimental validation of the built prototype have been carried out to verify its
operation. New solutions for the leg mechanism have been proposed.

1. Introduction
The walking in nature is a very flexible and complex task. For example, in
generating a trajectory several partshystems are involved: muscle as actuators,
bones as linkages, nerves as sensors, and brain as a complex control system [ 11.
Moreover, locomotion type depends on several variables. For example, one of
the most influent aspects can be considered the environment characteristic.
While the legged locomotion is more adaptable in a wide range of terrain, the
wheeled locomotion is faster but only on smooth surfaces. The most common
walking machines are wheeled and tracked systems, but large interest is also
focused on legged machines. A leg should generate an approximately straight-
line trajectory of a foot point with respect to the body, as outlined for example
in references [l-31. There are also robots whose design combines wheels, legs or
tracks. They can be divided in several categories: articulated legs having wheels
at their end, like for example in [4]; wheels and legs in separated modules, like
for example in [ 5 ] ; articulated tracks, like for example in [6];wheels and tracks
in separated modules, like for example in [7].
A combined design of legs and wheels can be a good compromise for a
walking machine combining the advantages of both locomotion types with the
aim to achieve a successful operation of mobile robotic systems in rough,
unstructured terrain, such as underground mines, forests, disaster sites, and
planetary explorations, as outlined in [8].

E-mail: tavolicri~~~iiiicas.it,
ottaviano@unicas.it, ceccarelli@unicas.it.

319
320

2. A 1-DOF leg for walking machine

The leg architecture is composed by a Chebyshev mechanism and a Hart
inversor, as shown in Fig. 1. The Chebyshev mechanism consists of the four-bar
approximate straight-line mechanism LEDCB. The lengths of the links are
determined in such a way that the shape of the point B trajectory is similar to the
one of a human step. Moreover, the straight part of the trajectory is quite
accurate. It is an important aspect in order to avoid the body lifting during the
walk. The Hart mechanism inverts and amplifies the trajectory. The
amplification factor, from point B to point A, is approximately 2. Thus, the body
can move horizontally by moving the feet and legs.
Several waking robots have a design with a pantograph as fundamental
mechanism and even Chebyshev linkage has been used for actuator purposes
like for example [%lo], specific experiences have been carried out at L A M in
Cassino, as reported in Ell-151.

(b)
Fig. 1. The new 1 -DOF hybrid robot: a) kinematic scheme for the I -DOF leg designed robot;
b) built prototype.
 321

3. The existing prototype and its operation

The mechanical design of the prototype has been developed by looking at
several aspects, such as
- low costs in terms of both manufacturing and energy
- easy-operation system in terms of control,
- a compromise between flexibility for walking tasks and simplicity of the
system.
The chosen material for building the robot is an aluminium alloy, because of
its characteristics of lightness, easy machining and good resistance to the
corrosion.
The designed robot consists of a biped module in which two legs are actuated
by a DC motor. Each leg has been provided by a suitable foot, which has been
rigidly installed at the leg tip, for walking in both flat and rough terrains. There
are two passive wheels, in order to have always three points in contact with the
ground.
Because of the use of only one actuator it is necessary to design a gear
transmission system, which transmits the actuation torque to both the legs. The
body of the mobile robot carries batteries for operating the robot. Furthermore,
it allows the installation of suitable sensors and acquisition system on board.
The walking synchronization is obtained by moving one leg in opposite phase
with respect to the other one. During the walking the prototype has always one
leg and two wheels in contact with the ground. The mass distribution and body
shape give that the projection of the barycentre is always in the area that is
limited by the elements in contact with the ground.
In order to improve the performances of the prototype several aspects have been
considered. The prototype has been tested indoor and outdoor over terrains of
various nature and slope. The prototype has been tested by walking forward and
backward on flat terrain. The prototype has been tested also for a slope of about
20 deg, and on outdoor various terrain, as reported in [16]. Figure 2 shows
examples of the walking tests. Parameters of the mechanism showed in Fig. 1 a)
are reported in Tables 1 and 2.

4. The attached problems

Main problematic aspects of the robot behaviour can be addressed to
- limited lift high,
- body shape and position of the barycentre,
- wheels and foots materials.
The built prototype has good flexibility, but one of its main limitation is the
little high of the lift. The idea of using actuated legs and passive wheels was
implemented in order to build a system able to climb obstacles of various size
and nature.
322

(4 co)
Fig.2 The built prototype while: a) climbing a slope of 20 deg; b) walking over unstructured terrain.

With the aim to improve this fundamental characteristic, and, at the same time,
to loose nothing in terms of easy-operability and low-cost, it is possible to
change the parameters of Chebyshev links, in order to have a higher trajectory,
allowing the prototype to overcome a larger obstacle variety, as reported in Fig.
3. In fact, the mechanism of the leg is the one which gives the trajectory size
and shape.
The above-mentioned problem can be formulated as:
r < sh
Oh 1 Sh
where r is the wheel radius, sh is the old lift high, Sh is the height of an obstacle
and Oh is the new step height. The new configuration allows the prototype to
overcome obstacles of Oh high, were Oh>sh. With the proposed configuration it
is possible to eliminate AAl link, having a linear inversion of the trajectory.
It is worth to underline that the backward walk has been obtained more
efficiently than in the case in pulling walk mode. The difference in walking
climbing directions can be explained by considering the position of the mass
centre and friction of the floor plane.
The velocity is constant during the walking on flat terrain, and it depends on
the friction between foot and ground. It is worth to underline that the backward
walk has been obtained more efficiently than in the case in pulling walk mode.
The difference in walking climbing directions can be explained by considering
the position of the mass centre and friction of the floor plane. In fact, because of
the actuator weight, the mass centre position is closer to the wheels and
therefore gravity helps the walking motion. This behaviour is emphasized when
the batteries are installed on board. In this case the weight of the system strongly
increases, and the position of the mass centre is even more decentred because of
the heaviest components lying on the passive module.
As further development of the prototype the body shape should be redesigned
with a rectangular profile. In this way the batteries can be the middle of the
body, improving both walking condition and stability.
 323

Fig.3 A scheme of the new configuration obtained by changing the Chebyshev mechanism: a higher
step is obtained (dotted line represents the new solution, continuous line represents the existing leg).

From a mechanical point of view it is possible to improve the prototype

walking by replacing the wheels with a version covered with rubber, in order to
avoid slipping on metallic surfaces. Another important design improvement can
be the replacement of the foot, with other one having spherical shape.

Table 1. Link parameters for the Chebyshev mechanism for the leg
design in Fig. 1 a).

Parameters (mm) a m C d f
Built Prototype 45 15 60 60 60

Table 2. Link parameters for the Hart mechanism for the leg design in
Fig.1 a).

Parameters (mm) p h z2 zj a z5 ~6 z7 zg zg
Builtprototype 168 104 225 150 150 225 225 150 225 100

5. A numerical simulation of the leg operation

A kinematic analysis has been developed in order to evaluate and simulate
performances and operations of the leg system. A fixed reference system CXY
has been considered attached at body in point C, as shown in Fig. 1. The position
of point B with respect to CXY frame can be evaluated as a function of the input
crank angle a and kinematic parameters of the Chebyshev mechanism LEBDC
in the form
x, = -a + m cos a + (c+ f)cos 8
(1)
Y, = -m sin a - (c + f)sin 8
The position of point A with respect to the fixed frame can be given as
XA = XM -(z3 + z,)cos (p, + (z6 + z8)cos (p6
(2)
yA= yM+ (z3 + z,)sin (p3 - (z6 + zs)sin (p6
324

The position of point AI with respect to the fixed frame can be given as

The transmission angles shown in Fig. 4.can be evaluated as

71 = - 0 + ( p 2 - K 7 y 2 = ( P 3 - ( p 6 --7c 973 =(P3 - ( P 2 + n (4)
The velocity of points B, A and Al can be evaluated by using time derivatives
from Eqs. (1) to (3).
The acceleration of point A, can be obtained with respect to the fixed frame as
XA, = ipi(z3 + z4) C O S C +
~ ~i/)3(z3+ ~ 4sin
) (p3 +
-@:(z(j + zg) cos (p6 -@6(z6 + 2 8 ) sin (p6 +

Figure 5 shows numerical results for the accelerations of the leg system when
the angular velocity o of the input crank is chosen with a constant value equal
to 1.0 rads.
By considering the obtained values of the main parameters (such as
accelerations and transmission angles) and the tested behaviour of the built
prototype new solution for the leg system can be considered.

k',(

Fig. 4. Numerical simulation for the transmission angles in degrees:

a) angle y,; b) angle y2; c) angle y,.
 325

0,c, -. . . . . . . ........ . . . . . . . . . . . . . . . . . . . . . . . . . . . . . ~ .............

(nl
82- .................................................................................. ..............

D?,- ................................... ........ ..,

101. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . ..........

Fig. 5. Numerical simulation for the accelerations: a) point B; b) point A; c) point A ,

6. Proposals for improving design the leg design

Design consideration and simulations have been developed in order to solve
the outlined problems. One of the main issues is related with walking versatility
for obtaining higher lift of the feet. A design attempt has been focused in
changing the link ratios with the aim to change the size of the coupler path in
the actuating Chebyshev mechanism.
By referring to the Chebyshev mechanism in the built prototype in Figures 1
and 6 a), the mechanism design is characterized by the ratios kl and kz that give
ED = CD = DB = k,LE
(6)
LC = k,LE
The convenient characteristics of the Chebyshev design can be preserved by
keeping the ratios in Eq. (6) and design improvements can be studied by
changing the ratio values. Among the many possibility we have considered
again Chebyshev mechanisms that are listed in [17].
Thus, two alternative solutions have been selected whose main characteristics
are reported in Table 3 as related with walking operation. Numerical simulation
have been carried out and results are reported in Figures 6 and 7, as compared
with the Chebyshev mechanism in the built prototype of Figures 1 and 2. In
particular, the size of the coupler path is different to give a higher lift of the feet
during the walking. Even the shape of the coupler is modified as shown in
Figures 6 b) and c), which will give a modified behaviour also for the Hart
mechanism that has been maintained unchanged. In fact, point M has a different
location and the Hart mechanism shows different configurations. Motion
characteristics are evaluated in term of transmission angles yI, yz, and y3 (see
Fig. 1 a)) in order to synthetically show the mechanical feasibility.
326

Table 3. Parameters for the leg mechanisms in Fig.6.

Leg Solution Leg in Fig. 6 a) Leg in Fig. 6 b) Leg in Fig. 6 c)

Lift (mm) 150 180 190

--.
kz 3.0 2.8 2.83
EDB (deg) 180 180 100
Min YI(deg) 48 77 37
Max YI ( d e d 91 153 108
Min ~2 (deg) 34 64 35
Max YZ (deg) 69 117 69
Min YJ (deg) 113 144 115
Max ~3 (deg) 142 166 148

. .
Fig. 6. Scheme and trajectories of the leg system: a) the existing leg; b) first proposal;
c) second proposal.
 327

(a) (b)
Fig. 7. Numerical simulation for the transmission angles y,, y2, y~ in degrees:
a) leg in Fig. 6 b); b) leg in Fig. 6 c),

In particular, the transmission angle yI refers to the force transmission

capability between the Chebyshev mechanism and Hart mechanism, as the case
of the leg functionality. Plots in Figures 4 and 6 are useful to check the
functionality of the proposed alternative solutions as compared with the original
one. In Table 3 numerical values summarize the results of numerical
simulations. One can note that although a considerable increase of the lift is
obtained, in general performances are not increased accordingly. In fact, the
range of the transmission angles is larger and the smooth behaviour, that is
indicated by the smoothness of the plots, is even deteriorated. Nevertheless, the
alternative proposed solutions can be considered feasible, since they refer to
successful practical mechanisms. In addition, they could be used as initial
designs in a design procedure which will use optimization techniques and
proper formulation for optimality criteria for walking mechanisms, as
concerning with the outlined problems in this paper.

7. Conclusion
In this paper a new prototype of leg-wheel walking robot is proposed with low-
cost easy-operation features. A suitable kinematic model has been presented for
a I-DOF leg with a fully-rotative actuation and for the whole system. A
kinematic analysis has been also carried out in order to evaluate and simulate
performance and operations of the walking robot. A low-cost leg-wheel
machine has been built at LARM in Cassino by using the proposed 1-DOF leg
mechanism. Experimental tests have been conducted over several terrains with
the aim of understand limits and possible improvements of the prototype. By
considering the leg structure two possible new solutions have been proposed in
order to have a higher lift.
328

References
1. A. Morecki, Biomechanical Modeling of Human Walking, World Congr. on
the Theory of Mach. and Mech., Vo1.3, pp. 2400-2403, Milan, 1995.
2. K. Yoneda, Design of Non-Bio-Mimetic Walker with Fewer Actuators, 4th
CLA WAR, pp. 115-126, Karlsruhe, 2001.
3. S.M. Song, J.K. Lee, K.J. Waldron, Motion Study of Two and Three
Dimensional Pantograph Mechanisms, Mech. and Mach. Theory, 1987.
4. G. Endo, S. Hirose, Study on Roller-Walker (multimode steering control
and self-contained locomotion), Procedings IEEE InternationaZ Conference
on Robotics and Automation, 2000, pp. 2808-2814.
5. S. Guccione, G. Muscato, The Wheeleg Robot, IEEE Robotics and
Automation Magazine, Vol. 10,2003, pp. 323-330.
6. F. Michaud, D. Letourneau, M. Arsenault, Y. Bergeron, R. Cadrin, F.
Gagnon, M.-A. Legault, M. Millette, J.-F. Pard, P. Lepage, Y. J. Morin,
Bisson, S. Caron, Multi-Modal locomotion Robotic Platform Using Leg-
Track-Wheel Articulations, Autonomous Robots, Vol. 18,2005, pp. 137-156.
7. C. T. Chen, Y.A. Hsieh, Mobile Robot, Us Patent 6.144.180,2000.
8. K. Iagnemma, S. Dubowsky, Mobile Robots in Rough Terrain, Springer
tracts in advanced robotics, Vol. 12, Berlin, 2004.
9. F. Pfeiffer, T. Zielinska: Walking: Biological and Technological Aspects,
International Centre for Mechanical Sciences Courses and Lectures n. 467,
Springer Wien New York, 2004, pp. 1-29.
10. M. H. Raibert, Legged Robots That Balance. MIT Press, 1986, pp. 3-14.
11. E. Ottaviano, M. Ceccarelli, C. Tavolieri, Kinematic and Dynamic Analyses
of a Pantograph-Leg for a Biped Walking Machine, Proceeding of 7th
CLA WAR, Madrid 2004, pp. 561-568.
12. C. Tavolieri, E. Ottaviano, M. Ceccarelli, A. Di Rienzo, Analysis and
Design of a 1-DOF Leg for Walking Machines, Proceedings of RAAD'06,
15th International Workshop on Robotics in Alpe-Adria-Danube Region,
Balantonfured, 2006, CD Proceedings.
13. C. Lanni, M. Ceccarelli, E. Ottaviano, G. Figliolini, Mechanisms for
Pantograph Legs: Structures and Characteristics, Proceeding of Xth
IFToMM World Congress, Oulu, 1999, Vo1.3, pp. 1196-1201.
14. C. Tavolieri, Design, Simulation and Construction of a Biped Wallung
Robot, Master Thesis, L A M , University of Cassino, Cassino, 2004.
15. A. Nardelli, Design and construction of a new wheel-leg mobile robot,
Master Thesis, LARM, University of Cassino, Cassino, 2006 (in Italian).
16. C. Tavolieri, E. Ottaviano, M. Ceccarelli, A. Nardelli, A Design of a Nex
Leg-Wheel Walking Robot, Proceedings of MED '07, 15th Mediterranean
Conference on Control and Automation, Athens, 2007, CD Proceedings.
17. I. I. Artobolevski, Mechanisms in Modern Engineering Design, Vol. 1-5,
Mir Publisher, Moscow, 1975-1980.
 Stiffness and Duty Factor Models
for the Design of Running Bipeds

Muhammad E. Abdallah and Kenneth J. Waldron

Mechanical Engineering Department,
Stanford University,
Stanford, CA, USA
(mesam, kwaldron}@stanford.edu

Supporting the design process for running biped robots, analytical models
are presented for two aspects of running: the duty factor (DF) of the gait, and
the stiffness value of the leg. For a given running speed, an optimal DF exists
that minimizes the energy expenditure. Based on a model of the energetics,
we present a formula for the optimal DF. This formula is validated by both
human data and simulation results. In addition, a model is presented for the
stiffness value of the leg as a function of the physical properties, speed, and
DF. The Gait Resonance Point is proposed as a design target for compliant
running. At this point, the gait matches the spring resonance and the stiffness
value becomes independent of the DF.

Keywords: biped robot, running, duty factor, SLIP stiffness, energetics.

1. Introduction
As the field of running robotics continues to progress, a more rigorous design
process is needed to meet performance criteria. This work addresses the
design of a running biped robot subject to gait specifications. We present
design rules and models for two aspects of running: the duty factor (DF)
of the gait and the stiffness value of the leg.
The presence of compliance in the leg of a running biped is invaluable
as an energy storage and thrust mechanism. It has been proven essential
to the running of biological systems.' It has also been instrumental in the
more dynamically successful robotic runner^.^-^
Surprisingly, little theory exists for selecting the stiffness according to
gait specifications. Raibert shares the effective stiffnesses for his hoppers,
but nothing on the selection process.2 Rad et al. designed their stiffness to
maximize the active energy input during stance.6 Schmeideler and Waldron

329
330

selected their stiffness value according to biomimetic model^.^ Ahmadi and

Buehler began with an analytical model of purely vertical hopping. They
then incorporated model-specific empirical formulas to relate the stiffness
to the speed.8 None of these studies provides a generalizable design process
or laws for selecting the stiffness of a general biped for any speed.
We present an analytical formula for the leg stiffness that is a function
of the physical properties, speed, and DF of the system. In addition, we
present the Gait Resonance Point (GRP) as a potential design target for
compliant running. At the GRP, the gait matches the spring resonance, and
the stiffness value becomes independent of the DF.
The DF is the fraction of a stride period a specific leg spends in contact
with the ground. It carries significant implications for the energy consump-
tion and ground impact forces of a gait. Despite these consequences, no
documented consideration for the DF, to our knowledge, has been given
for the design of any of the existing running bipeds. For a given speed,
an optimal DF exists that minimizes the energy expenditure of the gait.
Alexander shows the existence of this optimal DF, displaying the energy
cost for a human model running at two discrete speed^.^
Based on a model of the energetics, we present here a formula for the
optimal DF. The formula is applicable to any speed and design for the
biped, and it is validated by both human data and simulation results.

2. The SLIP Model for Stiffness Analysis

A simple, yet effective model for compliant running is the Spring Loaded
Inverted Pendulum (SLIP) model. It consists of a massless, spring-operated
leg attached to the center of mass of the body. It has been widely used in
the study of both biological and robotic systems, and its applicability has
been well documented.10-12Fig. 1 shows the model.
While running at steady-state, the SLIP model must exhibit a symmet-
ric contact phase.I2 This symmetry is reflected in both the configuration
and velocities, as shown in Fig. 1.
The gait consists of a contact phase followed by a flight phase. The
contact and flight times (t, and t f respectively) can be determined from
the kinematics of Fig. 1.
21, sin(e,)
t, = , t f = -2% (1)
VX 9
v, and v, represent the horizontal and vertical velocities at the initial mo-
ment of contact. 1, and 8, are the initial leg length and angle. g is the grav-
 33 1

- X 4-t
R
Fig. 1. The SLIP model. The free body diagram is shown to the left. The symmetry of
the contact phase is shown to the right.

itational acceleration. The calculation for the contact time approximates

the forward velocity as a constant vz.
McMahon and Cheng utilized this symmetry to analyze the dynamics of
running." They conducted numerical integration of the SLIP equations of
motion to determine the parameters that satisfy the described symmetry.
They presented empirical formulas that fit the data within errors of 0.5%.
Using their empirical formulas, we can display the spring natural fre-
quency of the system, un,as a function of the speed and DF. Fig. 2 shows
the results. The parameters used throughout this paper reflect an average
human: a mass of 73 kg, leg-length of 0.8 m, and a touchdown angle of 0.4
radians.
We will use McMahon's data as a benchmark for our approximations.
His formulas would suffice for determining the needed stiffness if it were
not for two factors. First, the formulas are very complex: 24 empirical pa-
rameters are needed to compute the stiffness at any one speed. Second, his
formulas are only valid for a limited range of conditions.
The next section will present an alternative analytical formula for the
stiffness and compare it to McMahon's results.

3. An Analytical Stiffness Model

To derive an analytical model for the SLIP, consider the dynamical equation
in the vertical direction. Fig. 1 shows the free-body diagram. k and I , are
the stiffness and initial length of the spring, and m is the mass.
332

The Natural Frequency for SLIP Running

26

S'

12
3 4 5 6 7 8 9 10
Speed (m/S)

Fig. 2. The natural frequency required for steady-state running of the SLIP model,
based on the empirical formulas of McMahon and Cheng." Each curve represents a
liftoff velocity (vz). The arrow points in the direction of increasing DF.

mz + + (mg- I C
ICZ ~ ,= o
cos(e)) (2)
Since the motion over the contact phase is known, we can replace cos(0)
with its average value. Denoted as z,
the geometric average is,
- sin(0,)
C$=-. (3)
$0

This results in a single-variable, linear equation of motion. This equa-

tion can be solved through boundary values based on the symmetry of the
contact phase.
We will introduce the contact frequency, w,, as the frequency of the
contact phase. Its period is twice the contact time.
IT
wc = - (4)
t C

The solution to this system gives us our final equation relating the
spring frequency to the system parameters. We will refer to it as the SLIP
Governing Equation. For brevity, cos(0,) is expressed in shorthand as c0,.

The stiffness a biped needs to run is a function of both physical prop-

erties and gait characteristics. It is a function of the mass, leg length, and
initial leg angle of the system. It is also a function of the desired speed and
 333

DF of the gait. The SLIP Governing Equation provides an analytical tool

for determining the stifhess value given those parameters.
The equation predicts the spring frequency with good accuracy. Com-
pared t o McMahon’s data, it produced errors less than 7%. A side-by-side
comparison is shown in Fig. 3.

Results of the SLIP Governing Equation

25

20

15

-
,....
SLIP governing equation
McMahon’s formulas
5t
3c l o

“ 4 5 6 7 8 9
Speed (m/s)

Fig. 3. The SLIP Governing Equation accurately predicts the spring frequency needed.
The plots compare results from the governing equation with McMahon’s empirical data
at different DF’s. The errors are all less than 7%.

4. The Gait Resonance Point

The SLIP stiffness curves display an interesting behavior: the stiffness val-
ues tend to converge a t a select speed. This phenomenon is exhibited in
Fig. 2. Although the stiffness is normally a function of both the speed and
the DF, at this speed the value is effectively constant, independent of the
DF. We call this point the Gait Resonance Point (GRP) for reasons that
will become apparent.
This behavior consistently demonstrated itself in McMahon’s formulas,
even with changes to the parameters. The SLIP Governing Equation (5)
allows us to solve for the point. The GRP represents the solution of the
equation at its singularity, where the tan term approaches infinity and its
coefficient approaches zero. This results in the following two conditions.
334

According to the first condition, the solution occurs when the contact
frequency matches the spring natural frequency-hence naming it the Gait
Resonance Point. These simple design rules predict the GRP with good
accuracy. Fig. 4 shows the analytical GRP alongside the frequency curves.

The Gait Resonance Point

speed (nlk)

Fig. 4. The Gait Resonance Point targets the intersection point, where the stiffness
becomes effectively independent of the DF. Shown as the star, the design rules predict
the point with good accuracy. These frequency curves are the same curves from Fig. 2.

We propose the GRP as a potential design target for compliant running

due to two advantages. First, the stiffness is effectively independent of the
DF; hence, we can implement different DF’s without changing the stiffness.
Second, the stiffness is governed by simple design rules, where the spring
frequency equals the contact frequency.
If the G R P does not meet the desired design specifications, one can
return to the S L I P Governing Equation to determine the needed stiffness.
I n this case, a D F needs to be selected. The next sections turn to the issue
of the DF. First, the energy cost is defined. Then, an optimal DF model is
presented.

5 . The Energy Cost of Running

5.1. Defining the Energy Cost
The kinetic energy (KE) of a multi-body system can be partitioned into
two segments. The first is defined as the external work, and it consists of
the KE of the center of mass (CM) in a ground reference frame. The second
is defined as the internal work,and it consists of the KE of the limbs in the
CM reference frame.l3?l4
 335

This framework allows for a simple characterization of the energy expen-

diture. By contrasting the top-of-flight and the bottom-of-contact instants,
the external work itself reveals two components. First, it must provide the
change in height, i.e. the hopping height energy. Second, it must provide
the change in KE as the system accelerates from its minimum horizontal
speed at the nadir to its maximum horizontal speed at the apex.
Similarly, the internal work can be characterized by two components.
First, it must provide the energy to swing the non-support leg, work con-
ducted primarily by the respective hip. Second, it must provide the work
conducted by all the other joints, primarily to maintain posture.
Of these four total components, two dominate the energy cost: the hop-
ping height energy and the swing energy. Each scales by the square of the
DF. We will thus consider these two only.
These two functions capture the tradeoff inherent in the DF. For exam-
ple, a smaller DF translates into a longer flight time. This long flight time
entails a large hopping height energy. At the same time, it increases the
stride period and thus reduces the swing energy required.

5.2. Deriving the Energy Cost

For the following energy calculations, we will consider the net absolute work
performed over a step. A step includes a contact phase and subsequent flight
phase, i.e. half the stride.
Given a change in height of h, the net absolute work performed for the
hopping height is 2mgh. Neglecting the change in height during contact,
the hopping height energy as a function of v, follows.
2
Ehh = mV, (8)
The swing energy, E,,, is derived e1~ewhere.l~
The resulting energy is,
E,, = 8,2m&(w2 - w:). (9)
1, and ml are the effective length and mass of the swing leg. w is the stride
frequency and w, here is the pendulum natural frequency. The total energy
cost, Etot, is the sum of both components.
The DF, P, is defined as the ratio of a leg’s contact time to the stride
period. Assuming symmetric right and left steps,

Given (10) and (l),we can solve for the total energy cost as a function
of the DF. This gives us our final energy cost formula.
336

where ci are constant functions of the parameters.

2
cg = 0,mll,g. (14)
This formula quantifies the energy tradeoff associated with the DF. Fig.
5 displays the energy cost for an average human running at 4 m/s.

50 -

Fig. 5. Plot of the energy cost formula, shown for an average human running at 4 m/s.
The energy cost at higher DF’s is dominated by the swing energy; while at lower DF’s
it is dominated by the hopping height energy. This tradeoff results in a global minima
that depends on both the running speed and physical parameters.

6 . The Optimal DF
To solve for the minima of the energy cost, we will optimize (11). This gives
us our final characteristic equation, the Optimal D F Formula.

(2)+ p4 2p- 1 = o (15)

This formula was validated in two ways. First, it was compared to data
of human running. As a widely accepted premise, biological systems adapt
 337

their gaits to minimize energy c o n s ~ m p t i o n . ~Minetti

> ' ~ experimentally de-
termined the DF of humans running at a range of speeds, using an average
sample size of over 30 runners.14
Second, the results were validated in simulation. Using the dynamic
model described in Appendix A, the biped was tested at two select speeds.
For each speed, a range of DF's was applied and the energy consumption
tabulated. The model approximated an average human running at speeds
of 3.5 m/s and 4.0 m/s respectively.
Both the human data and the simulation model exhibited close corre-
spondence with the Optimal DF Formula. Results of the formula nearly
fall within a standard deviation of the average human DF's. This close cor-
respondence occurs despite neglecting the error between mechanical and
metabolic work, and not knowing the actual physical properties of the run-
ners. Fig. 6 displays the combined results, comparing the DF's from the
formula with the human data and simulation results. The Optimal DF
Formula is computed here with the same average human parameters used
throughout.

Comparison of DF Results

0 05 1 15 2 25 3 35 4

Speed ( d s )
Fig. 6. The Optimal DF Formula closely corresponds to both simulation results and
natural human DF's. The experimental human data is provided by Minetti and is dis-
played as a mean f S.D.14
338

7. Conclusion
This work answers the following basic question: I want to design a biped
robot to run at a target speed. What DF should I design it for? And what
stiffness value does it need?
In designing a running biped, selecting a stiffness to meet a desired gait
specification is not trivial. It depends on the speed, configuration, and DF
in highly-nonlinear systems. The DF itself warrants careful consideration,
given its effect on the energy consumption and impact forces of the gait.
It is one thing to design a compliant robot that can run stably; it is
another thing to design it for set gait specifications. This work presents
formulas and design laws that supplement the design process for a running
biped robot.

Appendix A. The Simulation Model

The simulation system modeled a telescoping legged biped, shown in Fig.
A1. The model consists of three rigid bodies, representing the torso and
two legs, connected by a revolute joint at the hip. The feet are point-masses
connected by prismatic joints to the legs.

Fig. A l . The biped model used for the dynamic simulation.

A fully dynamic simulation was created. The foot-ground contact was

modeled as a rigid, inelastic contact using motion constraints. Friction was
assumed sufficient to avoid slip. The model implemented the control strat-
egy of Abdallah and Waldron for running.17 It's parameters modeled an
average human as shown in Table A l . A detailed description of the model-
ing and experiment results is available e1~ewhere.l~
 339

Table Al. Model Parameters

I distance from mass inertia
I hip t o CM (m) (kg) (kg m2)
torso I 0.34 48.3 8.12
0.384 11.5 1.03
0.8 (nominal) 1.0 0

References
1. R. Alexander, Elastic Mechanisms in Animal Movement (Cambridge Univer-
sity Press, Cambridge, 1988).
2. M. Raibert, Legged Robots that Balance (MIT Press, Cambridge, MA, 1986).
3. J. Nichol, L. Palmer and K. Waldron, Design of a leg system for quadraped
gallop, in Proceedings of the 11th Congress in Mechanism and Machine Sci-
ence, ed. T. Huang (China Machinery Press, Tianjin, China, August 2003).
4. M. Ahmadi and M. Buehler, The ARL Monopod I1 running robot: control and
energetics, in I E E E International Conference on Robotics and Automation,
(Detroit, MI, 1999).
5. B. Brown and G. Zeglin, The bow leg hopping robot, in I E E E International
Conference on Robotics and Automation, (Leuven, Belgium, 1998).
6. H. Rad, P. Gregorio and M. Buehler, Design, modeling and control of a
hopping robot, in IEEE/RSJ Conference on Intelligent Systems and Robots,
(Yokohama, Japan, 1993).
7. J. P. Schmiedeler and K. J. Waldron, Leg stiffness and articulated leg design
for dynamic locomotion, in Design Engineering Techinal Conferences and
Computers and Information in Engineering Conference, (Montreal, Canada,
2002).
8. M. Ahmadi and M. Buehler, I E E E Transactions on Robotics and Automation
13,96(February 1997).
9. R. Alexander, Phil. Trans. of the Royal Society of London 338,189 (1992).
10. R. Blickhan, Journal of Biomechanics 22,1217 (1989).
11. T. A. McMahon and G. C. Cheng, Journal of Biomechanics 23,65 (1990).
12. W. Schwind, Spring loaded inverted pendulum running: a plant model, PhD
thesis, University of Michigan, (Ann Arbor, MI, 1998).
13. G. Cavagna and M. Kaneko, Journal Physiology 268,467 (1977).
14. A. E. Minetti, Journal of Biomechanics 31,463 (1998).
15. M. Abdallah, Mechanics motivated control and design of biped running, PhD
thesis, Stanford University, (Stanford, CA, 2007).
16. D. Hoyt and R. Taylor, Nature 292,239(July 1981).
17. M. Abdallah and K. Waldron, A physical model and control strategy for biped
running, in I E E E International Conference on Robotics and Automation,
(Rome, Italy, 2007).
 Constraint Based Trajectory Simplification of f i l l Body
Trajectories for a Walking Robot

Hanns W. Tappeiner, Alfred A. Rizzi

The Robotics Institute, Curnegie Mellon University,
Pattsburgh, PA 15213, USA
E-mail: hannsOri. crnu.edu
h t t p : / / ~ . ri.emu. ~u/people/tappeiner-hanns.html

Keywords: Walking; Legged Robots; Trajectory Simplification;

1. Introduction
In this paper we present an algorithm and experimental results on a real
robot which solves two related problems: First we simplify a robot trajec-
tory created by a human driver by reducing the amount of control points,
according to a set of constraints. Second - as a side effect - we eliminate
local small scale driver errors. The goal is to generate trajectories which
axe closer to what the driver “intended” to do and can be used as good
examples for robot learning strategies.

Fig. 1. A) Robot LittleDog climbing over a terrain board. B) An accurate 30 model

of both LittleDog and terrain is used to perform full body collision detection, COM
checks etc. The arrows mark collision points, the orange line is the robots COM position
trajectory over the terrain.

In our experiments with controlling a four legged walking robot over

340
 341

rough terrain, it turns out that a skilled human driver, equipped with a
good control device like a joystick or gamepad, can outperform our used au-
tonomous control techniques. Even though controlling a 12dof, four legged
robot with a joystick is not straightforward, it seems that the driver devel-
opes a level of understanding of the terrain and the corresponding control
actions of the robot which is more effective than our autonomous control
techniques.
It seems reasonable to consider using examples from a human driver to
learn autonomous behavior for the robot. One problem is that a driver
does not create “clean” trajectories but often introduces irrelevant motion
or even makes mistakes, especially at the small scale: for example the op-
erator may try to place a leg several times before actually ending up in a
good spot. Additionally it is difficult for an operator to control multiple
axes at the same time with a joystick, and the result is somewhat jerky mo-
tion of the robot. Even though the overall trajectory is probably very good
and captures the idea the driver had to solve the problem, it will contain
a lot of unneccessary noise at the smaller scale. We would like to remove
that “noise” and produce trajectories which are closer to what the driver
intended to do rather then what he actually commanded.

Fig. 2. Example trajectory and the corresponding simplified trajectory. The red x in-
dicate constraint violations, in this case simply unwanted collisions with the terrain.

Secondary, our recordings of trajectories generated by a human driver

contain control points at every timestep, which means a large amount of
data. It is possible that there are only very few key control points in the
342

motion which need to be there and everything else is much less important.
Reducing the number of control points may make trajectory learning more
effective.

2. The basic algoritm

Figure 2 shows four subsequent snapshots of the algorithm simplifying a
trajectory. Figure 3 the result of applying the algorithm to a real robot
trajectory. We start with a trajectory, containing control points at every
timestep. In addition, we define a set of constraints. A robot state s, at a
certain point in time t, is valid if all the constraints are fullfilled. We could
remove contol points from the original trajectory and interpolate between
the remaining points until we violate some of the constraints.

Fig. 3. Closeup which shows operator introduced noise in the trajectory and the cor-
responding simplified version.

In this version, we do the inverse. We start with only two control points
in the trajectory, the first and the last, interpolate between them and check
the constraints. If we violate some of them, we add an additional point from
the original trajectory in between and check the constraints again. We keep
doing this until the whole trajectory is valid. As we will see later, the real
application requires a set of more complex constraints (approximate colli-
sions, valid center of mass) than 2D collisions to generate valid trajectories.
 343

3. Test Results
In our experiments, a full body state of the robot is defined by 19 parame-
ters: Position (xlylz) and orientation (quaternion) of the body in the world
and the 12 joint angles of the legs.

The algorithm adds control points S, to the simplified trajectory and

interpolates between them until no constraints are violated. There are two
main groups of constraints: collisions with the terrain and keeping the cen-
ter of mass (COM) above the support triangle. The algorithm has to keep
existing wanted collisions (like footsteps) but not introduce unwanted col-
lisions. The simplified trajectory is valid if:
(1) It produces the same collisions (time 7 and position w) with the envi-
ronment as the original: 1 1 -~T~11~ < E t l I/wo- w,I1 < e p
(2) It has a COM above the support triangle of the stance feet at any time
the original has one.

Fig. 4. 3D model of TerrainG with the four tested walking directions GO, G1, G2 and
G3.

Depending on the complexity of the terrain the algorithm is able to re-

duce the amount of control points from about 92% (complex rock terrain)
to 99.7% (flat ground). Figure 4 shows one of the terrain boards from the
344

LittleDog program, TerrainG. For the two directions GO and G2 a human

operator created a trajectory that reliably crossed the terrain at different
speeds. Those trajectories have been simplified according to the algorithm
described in this paper, and original and simplified version have been com-
pared in terms of behavior on the real robot and number of remaining
control points. Table 1 shows the result of this comparison.

Table 1. Comparison of original and simplified trajectory properties on TerrainG.

Trajectory Control Points Fastest run [sec] Average joint speed [rad/sec]
Original GO 21240 21 (lox) 0.0758
Simplified GO 301 (1.4%) 15 (14x) 0.0624
Original G2 27417 28 (lox) 0.0825
Simplified G2 417 (1.5%) 19 (14.7~) 0.0702

The amount of control points is usually reduced by about 97% and the
maximum overall speed is higher for the simplified trajectory because of
smoother body and leg motions. Also, even though the maximum forward
speed is higher, the average joint speed during the runs is lower in the
simplified trajectories, because the simplification cancelles out substancial
parts of unnecessary motions (parts of noise and driver mistakes).
Figure 5 shows views of a full body trajectory for each leg and robot
body (z coordinate over time) on a rock board. The original trajectory
contains 12780 control points, the corresponding simplified trajectory 47
control points, which is 0.37% of the original. This plot shows 2D views of
the high dimentional trajectory.

Videos of the simplification process, the robot crossing TerrainG using

original and simplified trajectories etc. can be found online at:
http://www.cs.cmu.edu/-hanns/clawar07/index.htm

4. Summary
In this paper we developed a simple but effective algorithm to reduce the
amount of control points in legged robot trajectories according to a set of
defined constraints. The effect is that the simplified trajectories are cleaner
in terms of noise and small scale driver errors than the originals and are
likely to be closer to what the driver intended to do. We think that such
trajectories will be more useful1 as examples for learning algorithms and be
better trajectories for trajectory libraries [2]. The results of the simplifica-
 345

-0.05
180 200 220 240 260 280

.-
180 2w 220 240 280 280
tlnl?

o.o;,.
-0.05
.-.- -4 4
? a
' \
-* \.* . t-t.. .* c . '
,
6
F - *.r.

-01
-0.15
180 200 220 240 260 280 '
O1
0.
r*-., *- -. . . ...... .
*
t. -4
b.~\
'
-i
* .
. c-+-*.. . .
I --- c --
time

-0.1
180 200 220 240 2M) 280
0.22 time?
0.2
0.18
0 16 --I--
r C _ _ *---+t. t. *-+** *** -.--- c- -c* .-.
0.14
0.12
160 2w 220 240 260 280
timi

Fig. 5. Z coordinate over time views of a trajectory for all four legs (FL, FR, HL, HR)
and body, both original and simplified.

tion have been tested on the real robot as well as in simulation and have
been compared to original trajectories on the robot. In all experiments, the
described algorithm produced trajectories which contain only a fraction of
the control points of the original trajectories. The simplified versions are
not only executable on the robot, but allow the robot to move at a faster
overall speed with smoother motions.

5. Future work
There are three main improvements to the algorithm which we think would
be very useful. The first one is that the actual constraints don't take into
account robot dynamics. This turns out to be ok at most of the speeds we
are walking at so far, but as the robot becomes faster, dynamics plays a
more important role and should be part of the constraints and the point
selection algorithm.
346

Even though the reduction of control points by almost always more than
95% is quite impressive, the amount of remaining control points per step
varies between 3 and 30, depending on the terrain etc (there where a few
outliers with about 90 control points per step). In some cases a different
point selection algorithm would produce better results, e.g. by not picking
the point in the middle, but near the actual collision. Instead of using linear
interpolation, we plan to fit splines trough the remaining control points to
get even smoother trajectories.
We also plan to use this algorithm only as a first step to guess the amount
of control points needed to generate valid trajectories. In a second step, we
would like to optimize the positions of those control points, for example to
maximize the robustness of the trajectory.

6. Acknowledgements
The LittleDog project is supported by the Defense Advanced Research
Projects Agency (DARPA) and part of the DARPA Learning Locomotion
Project.

7. References
References
1. Footstep Planning for the Honda ASIMO Humanoid - J. Chestnutt, M. Lau,
K.M. Cheung, J. Kuffner, J.K. Hodgins, and T. Kanade (Proceedings of the
IEEE International Conference on Robotics and Automation, April, 2005).
2. Policies based on trajectory libraries. Martin Stolle and Christopher G. Atke-
son (Proceedings of the International Conference on Robotics and Automation
(ICRA), 2006).
3. A concept for manipulator trajectory planning - F. Pfeiffer and R. Jahnanni
(Proceedings of the IEEE International Conference on Robotics and Automa-
tion, pages 1399-1405, 1986).
4. An integrated method for planning smooth collisionfree trajectories for robot
arms - J. Zhang (Proceedings of Third Annual Conference on AI, Simulation
and Planning in High Autonomy Systems, Perth, pages 46-53, 1992).
5. Case Studies in Trajectory optimization: Trains, Planes, and other Pastimes -
Robert J . Vanderbei (Operations Research and Financial Engineering Prince-
ton University, 2000).
 FOOT PLANNING MOTION OF HUMANOID ROBOT RH-1
USING LAG ALGORITHM*

MARIO ARBULU, LUIS CABAS, DMITRY KAYNOV, PAVEL STAROVEROV,

CARLOS BALAGUER~
Department of Systems and Automation Engineering, University Carlos I l l of Madrid,
Av. de la Universidad 30
Leganks, Madrid 28911, Spain

Real-time gait generation algorithm for humanoid robot is presented in this paper. The
3D foot motion planning for the humanoid global motion is developed in order to walk in
any surface as plane, ramp, climbing stairs. Furthermore, it is possible continuous
change the step length and orientation in real time.. The proposed algorithm takes into
account physical robot constraints as joint angles, angular velocity and torques. Torques
are computed by Lagrange method under Lie groups. Some results are shown and
discussed..

1. Introduction
The gait pattern generation of humanoid robots has several solutions and it is a
no closed problem, so we propose in this work a feasible solution of foot
planning for being implemented in real time. Two reference trajectories should
be introduced to the humanoid: The COM trajectory and swing foot trajectory.
The COM trajectory [ 5 ] , [6] must maintain the robot stability while it is walking.
The swing foot trajectory [l], [3] should change the humanoid position avoiding
support leg collision, obstacle collision, maintain stability limit (ZMP in stable
zone [2], [4]) and adapt on landing surface, reduce the reaction force on the floor
when it is landing, and reduce the joints angular velocity in particular of knee
joint. In this work, foot planning motion of Rh-1 humanoid robot are described
in order to get stable walking, and decide in real time the foot placement, so the
“Local Axis Gait” (LAG) algorithm is proposed. The performance of this
algorithm is described as following: at first, goal foot position and orientation
should be decided; it is composed by 3 parameters at local suitable axis: step
length (LJ, lateral foot motion (L,) and final foot orientation (Le). After that, a
5‘h order polynomial interpolator is proposed for going from actual foot position

This work is supported by CICYT (Comisi6n Interministerial de Ciencia y Tecnologia).

347
348

to goal foot point at local axis, which allows us to control the position, velocity
and acceleration of the foot motion. Human like foot motion is obtained and
smooth joints motion could reduce the mechanical efforts. For avoiding support
leg collision a preview motion range study of swing foot had been done.
Furthermore, for avoiding obstacle collision, maintain stability limit and adapt
on landing surface, robot sensors decide the goal foot position and orientation.
The present work is divided as following: at first, in section 11, an overview of
foot motion planning is developed. In section 111, motion interpolators are
developed, and planning foot position and orientation are introduced. In the
section IV, dynamics analysis is detailed in order to obtain the torque limits on
each robot joint by using Lagrange formulation, Lie groups and screw theory. In
section V, some simulation on Rh-1 simulator are shown and discussed. Finally
in section VI the conclusions of this work are presented.

2. Overview of foot motion planning

Figure 1. Foot motion planning, at local axis inp" (support foot) for doing n-th step

The foot motion planning for doing the n-th step should be modeled, as shown
in Fig. 1. The swing foot moves from actual configuration (position and
orientation (n-I)-th frame) to the goal configuration (position and orientation
(n+l)-thframe).
The input parameters are the swing foot goal configuration (position and
orientation), defined in appropriate local axis as: (Lr, Ly, Lg)"*'. It is possible to
identify for the n-th step the parameters and frames as following:
 349

c : World frame

c , C n-l , c n+l : n, n-1, n + l local configuration frame, support foot

pn-l P"+l
Pn : n, n-1, n + l position of support foot local frame

: goal swing foot motion and orientation

Goal foot configuration (position, p"", and orientation, L T ' ) are the input
parameters for doing the next step. It could been obtained by humanoid sensors
or external command. Those input parameters could be generalized in order to
compute the n-th step in real time as:

As see in (l), goal foot configuration depends of support foot position,

because the support foot is the local axis of gait input parameters.

3. Swing foot trajectory

Figure 2. . Swing foot motion planning changing walking direction in real time. The n-fh step
trajectory in black bold line, and swing foot is represented by the red bold rectangles, while the
support foot is represented by the blue bold rectangle

Polynomials have been used for approximation because they can be evaluated,
differentiated, and integrated easily and infinitely many steps using just basic
arithmetic operations of addition, subtraction and multiplication (2). A
polynomial of order n or degree n is a function of the form:
350

@(t) = w1 + w2 .t + ... w, .t n-1 = + cn

k=l
W k .t
k -1

Where @ ( t ) E p n ,being P" the set or linear space of all polynomials of

order n .
The swing foot motion is planned by two fifth order interpolators (3), one for
climbing the foot (+I@)) and one for landing the foot (+&)), on each axis motion
and orientation motion.
Those ones allow to control the foot position, velocity and acceleration, so it
is minimized the foot landing reacting force, the knee angular velocity, smooth
and natural joints motion is obtained.

2
( t )= W: .t5 + wk .t4 + wi.t3 + w:.t2 + wi.t + W: (3)
d t )= T
@2 (t,x i ' , i:,S:, x;, x;, S;)2 V - I t 5T
2
2 5 2 4 2 3 2 2 2 2
$b2(t)=W6.l + w g 1 + W 4 J +w3.t +w2.t+w1

Where the fifth order polynomial are characterized by the next boundary
conditions on each zone; i.e. for j-th zone O=lfor climbing, j=2 for landing), as it
have seen in Fig. 2:

T : Step time
j .j
( x i ,xi ,xi
..j
) : Initial position, velocity, acceleration of j-th zone
. . .
( x J ,x J ,X J ) : Final position, velocity, acceleration, ofj-th zone
f f f
So the polynomial coefficients could be developed as following from
boundary conditions described above:
 351

The continuity at the ending of climbing zone and at the starting the landing
one, should satisfied the next conditions:

& ( T / 2 )= & ( T / 2 )
h ( T / 2 )= & ( T / 2 )
4 ( T / 2 )= 4 ( T I 2 )

So, the a set of six splines for swing foot motion are obtained, three for spatial
motion and three for spatial orientation. Furthermore, natural and smooth foot
motion could be developed at low computational cost for being applying in real
time.

4. Dynamic Modelling

4.1. Lugrange formulation

In order to compute the joint torques and dynamics constraints, dynamic model
should be proposed, so Lagrange formulation under lie groups and screw theory
had been developed [7]; because that give us a natural description of jacobian
manipulator.
The Lagrange formulation could be expressed as following (Lagrangian):
352

Being the motion equation by r torques the next:

Which could be expressed as following:

Mm(8).8 + N ( 0 , b ) = r (13)
So, Mm(8)E %nxn is the inertial manipulator matrix and it is defined by
the operational jacobian manipulator .Isliand the inertial generalized matrix M ias
(14):
n
Mm(8) = J$i(8).Mi.J,,j(8) (14)
I
The Rh-1 humanoid robot has 21 degrees of freedom, for this fact
~ 21x21 .
M m ( 8 )%
The N ( 8 , b) term give us the joints torque due to the potential forces, and it
is defined as (15):

n
From (15), v(e)= z m i . g . h i ( 8,)where miis the mass of the i-th link, g is
i=l
the gravity acceleration and hi(B) the COM height of the i-th link.

5. Simulation results

5.1. Rh-1 humanoid robotplatform

Several simulation tests have been done in Rh-1 simulator software,
developed under VRML environment, before to test in Rh-1 humanoid robot
platform (Fig. 3), which specifications are described in Table I.
 353

(a) ro)
Fig. 3. Rh-1 humanoid robot (a) and its mechanical configuration (b)

Table 1. Rh-1 humanoid robot specifications

Link Number Of Total Joint Human angular range Rh-1 angular range
DOF
Head Head -50 * to 50" (Roll) -
-50 'to 60" (Pitch) -
-70 to 70" (Yaw) -
Waist 1 (Yaw) 1 Waist -50" to 50" (Roll) (Roll)
-30"to 45' (pitch) - (Pitch)
-40"to 40"(Yaw) -45" to 45" (Yaw)
Arm 4 4x2 Arm
1 (Pitch) Shoulder -90"to 0" (Roll) -95" to 0" (Roll)
Shoulder 1 (Roll) -180" to 50" (Pitch) -180"to 60"(Pitch)
-90" to 90" (Yaw) -
1 (Pitch) Elbow -145" to 90" (Pitch) -135"to 0" (Pitch)
Elbow -90"to 90" (Yaw) -
Wrist 1 (Roll) Wrist -55" to 25"(Roll) -90" to 90" (Roll)
-70" to 90" (Pitch) -
k g 6 6x2 Leg
Hip l(Yaw) Hip -45"to 20"(Roll) -20"to 20"(Roll)
1 (Roll) -125"to 15" (pitch) -80" to 80" (Pitch)
1 (Pitch) -45" to 45"(Yaw) -15"to 15"(Yaw)
Knee 1 (Pitch) 1x 2 Knee 0" to 130"(Pitch) 0" to 100" (Pitch)
Ankle 1 (Roll) 2 x2 Ankle -20"to 30" (Roll) -20"to 20"(Roll)

5.2. Gait patterns and constraints

Some gait patterns should be seen in Fig. 4, taking into account the angular joint
354

range limits as see in Table 1. The proposed gait generation algorithm allows to
change directions while the humanoid is walking, continuous step length
changing and walking in any surface.
Angular velocity constraints have been satisfied as actuator limits, thus the
smooth gait patterns could be validated; i.e the actuator knee angular velocity
limit is 8000 RPM, the gait pattern generated is suitable for this requirement as
we can see in Fig. 5. The next two joints (Fig. 6) drive the sagital and frontal
ankle support leg motion, 2 and 1. The joints actuator torque limits are defined
by the reduction gear and motor nominal torque. So, the limit nominal torque for
actuators 1 1 is 200 N.m, as it is shown in the chart that each joint never overlap
the nominal torque limit; it is notice that the knee joint have the highest torque,
the gait pattern generated should have in account to reduce it. The 1 joint torque
limit is 17.6 Nm, the chart shown 12 N.m as maximum, it is near the nominal
one, because the frontal ankle joint has less reduction respect to other ones.
Furthermore, this joint support overall humanoid robot mass during single
support stage and it drives the frontal balance in order to maintain the ZMP
inside the convex hull.

Fig. 4. Gait patterns generated for continuous humanoid walking. It is possible to'generate any
direction continuous walking

Fig. 5 'Ihe angular velocity requirement by knee joint, its actuator run up 8000 RPM
 355

Fig. 6. AnMe joints torque of the support foot in the n-th step

Fig. 7. Snapshots of VRML Rh-1 simulation, for smooth naturally and continuous walking as
Figure 4.

5.3. VRML Rh-1 s i m u l a ~ n s

Snapshots for doing a circle and curve path on flat surface have shown in Fig. 7.
Natural and smooth change direction could be done by the proposed algorithm,
thus natural walking is obtained as human one. Furthermore, faster walking
motion should be developed until 1 Km/h at any direction on flat surface; in
another kind of surface slower walking motion by physical constraints and
stability one would be obtained.

6. Conclusion
Foot motion planning had been developed for applying in real time, by satisfying
the stability criterion on bipedal walking robots, joint angle, angular velocity and
356

torque conditions.
Lagrange formulation, Lie groups and screw theory had been employed in
order to verify the joint torques in the humanoid robot, while it is walking, thus
dynamics constraints are obtained.
Simulation results on Rh-1 simulator environment had been shown, and it
allows to test the gait patterns before to apply it in actual humanoid robot. The
next step is to test the above algorithm in actual Rh-1 humanoid robot platform.

Acknowledgments
We would like to thank the members of the Robotics Lab of University Carlos
111 of Madrid for their cooperation an suggestions. We also express
acknowledgment to the Humanoid team for their support.

References
1. H.Lim, Y.Kaneshima, A.Takanishi, et al., “Online Walkinge Pattern
Generation for Biped Humanoid Robot with Trunk,” Proc. of IEEE Int.
Conf. on Robotics and Automation, pp.3111-pp.3116, 2002
2. M. Vukobratovic and A. Frank, “One the Gait Stability of Biped Machines”,
IEEE Transactions on Automatic Control, Dec. 1970.
3. M. Morisawa, K. Harada, S. Kajita, F. Kanehiro, K. Kaneko, K. Fujiwara,
S. Nakaoka and H. Hirukawa, “A Biped Pattern Generation Allowing
Immediate Modifications of Foot Pacement in Real-time”, Humanoids ’06,
Dec. 2006.
4. K. Hirai, M. Hirose, Y. Hikawa and T. Takanaka, “The development of
Honda humanoid robot”, IEEE International Conference on Robotics and
Automation (ICRA 1998) Leuven (Belgium)
5 . Kajita, S., Kanehiro, F., Kaneko, K., Fujiwara, K., Harada, K., Yokoi, K.
and Hirukawa, H., “Biped Walking Pattern Generation by using Preview
Control of Zero-Moment Point,” Proc. of ICRA2003, pp.1620-1626,2003.
6 . M.Arbulli; L.M.Cabas; P.Staroverov; D.Kaynov; C.PBrez; C.Balaguer. “On-
line walking patterns generation for Rh-I Humanoid Robot using a simple
three-dimensional inverted pendulum model”. 9th Internacional Conference
on Climbing and Walking Robots (Clawar 2006). Brussels. Belgium. Sep,
2006.
7. R.M. Murray, Z. Li, and S.S. Sastry, “A Mathematical Introduction to Robot
Manipulation”, CRC Press, 1993
 GAIN PROPERTY FOR BIPED WALKING VIA LEG
LENGTH VARIATION

Tetsuya Kinugasatt, Shoichi Miwat, Yannick Aoustint and Christine Chevallereauf

t Institute de Recherche en Communications et en Cybenzetique de Nantes (IRCCyN)
BP 92101, 1 rue de la Noe, 44381 Nantes cedex 03, FRANCE
$ Okayama University of Science
1-1, Ridai-cho, Okayama, 700-0005,JAPAN
{ Tetsuya. Kinugasa, Yannick.Aoustin, Christine. Cheualereau}@irccyn.ec-nantes.fr,
miwa@mech.ous. ac.jp

The purpose of this paper is t o show a gain property of frequency response

analysis for a biped based on passive dynamic walking mechanism. The walking
method is composed of a cyclic length variation of legs and a body constraint
control. Some simulated results show that the walking gain property represents
an efficiency in the sense of consumed energy. It was found from the result that
walking frequencies only depended on frequencies of the leg length variation,
and there is a resonance frequency which can be called a walking resonance
frequency. Finally, our biped called Prototype Biped Emu-IV (PBEmu-IV) is
introduced, and some experimented results are shown in order t o verify the
effectiveness of the frequency analysis.

Keywords: frequency analysis; passive dynamics based walking; variable length

leg.

1. Introduction
It has been reported that Passive Dynamic Walking (PDW)l by McGeer
is extremely natural and efficient. There was no reference trajectory and
walking is not constrained by ZMP. Almost of all basic ideas for actu-
ated walking based on PDW were given in his paper (e.g.2). Now many
researchers are trying to give improvements and to realize such kind of
walking by experiments with stability analysis (e.g. in the l i t e r a t ~ r e ~We
~).
are focus on walking by cyclic variation of leg length which is one of the
methods of McGeer. However the given analysis is only for stability of a
cyclic motion.
The objective of this paper is, therefore, t o analyze the frequency re-
sponse of biped walking by the sinusoidal leg length variation. The gain

357
358

property is defined and it gives the same efficiency of walking as consumed

energy. Finally, some simulations and experiments are shown to verify a
validity of the analysis.

2. The Biped Modeling

2.1. Dynamic model of single support phase
The biped appeared in Figure 1 is composed of a body and two symmetric
legs which consist of prismatic actuated knees and spherical feet. Feet radius
are chosen so that the biped can be originally stable for its upright position.
The hips are rotational actuated joints with a little viscosity friction. We
assume that there is no slip between the spherical foot and the ground. The
equation of motion is as follows:
+ +
De(qe)Qe {he(qe,4e) C}Oe + G e ( q e ) = Ber + DRl(qe)Rl, (1)
ji'H = -{(11- R ) Q ~ 2q1il)
- cosql - (i; + (z1 - R ) Q sinql
~ ) - R Q1 ~ (2)
~ -{(11
j j = - R)Q1+ 2q1il)sinql + (i; - (11 - R)qt)cosql. (3)
qe = [qi, qz,q3,11,/2, X H , Y H ] ' , and X H and YH are the hip position. The

Fig. 1. Biped model and Experimental apparatus PBEmu-IV

Table 1. Physical parameters for the dynamic model

m, l[kg] I, 0.05[kgm2] rS O[m] sh 0.4[m] L 0.78[m]

mf 1.5[kg] If 0.05[kgm2] ~f 0.15[m] fm 0.3[m] R 0.78[m]
mb 15[kg] Ib 3[kgm2] d O.o1[P,s] Sb O[m]

counterclockwise rotation is positive. D,, he,C E !R7x7,GeE !R7 are an

 359

inertial matrix, a Coriolis and centrifugal effect matrix, a viscosity effect

matrix and a gravity vector, respectively. Be,D R E~ S7" are a coefficient
matrix for the inputs and a Jacobian matrix connecting to the stance leg tip
and joints, respectively. r = [rl,rz,F 3 , r,]' consists of relative torques of
the first hip, the second hip, the first and the second knees. R1 = [R,1, R,1]'
represents ground reaction force vector at the stance leg tip.

2.2. Impact model

We assumed that the impact is inelastic and instantaneous without slip.
After the impact, the biped is sure to return the single support phase. The
following impact equation is given:
D e ( q e ) ( q z - )0, = DRZ(qe)Im2. (4)
The impact occurs when the leg tip of the swing leg contacts to the ground.
qe in equation 4 is the state vector at the impact. q, and q,f are the angular
velocities just before and just after the impact. D R is ~ a Jaxobian matrix
connecting to the swing leg tip and joints. Imz = [Im22,Imy2]' is an impulse
vector at the contact point of the swing leg.

3. Control method and Walking Gain

3.1. Sinusoidal leg length variation and body control
Sinusoidal functions with respect to time are used for reference trajectories
of leg length which are controlled by PD law with the gravity compensation.
Let K,, w, and 4, be reference amplitude, angular frequency and phase
angle, respectively, then the reference trajectory is described as follows:
lid = K , sin(wt - q5w + ri) + L , i = 0,1, (5)
i i d = K,w C O S ( W ~- &, + ri), i = O , 1 . (6)
The PD law is also used for attitude control of the body and realized by
the torque at the hip joint of the stance leg. The control laws are written
in the following equations:
rl = K p b q 3 -k K d b 4 3 + G c b ( q e ) , (7)
r2= 0, (8)
r3 = - K p k ( l l - l l d ) - Kd/c(il - i l d ) +Gcl(qe), (9)
r4 = - K p k ( h - h d ) - Kdk(i2 - izd) 4- G c z ( q e ) , (10)
where K p , , K d * and G,, are proportional gains, differential gains and the
gravity compensations, respectively.
360

3.2. Definition of walking gain

All variables are defined when the motion is cyclic. Let w be a angular
frequency for the sinusoidal input of the legs. Let one step be a motion
from just after the impact to just after the next impact, and n be the step
number. Let T, be a walking cycle[s] which is a duration for two steps.
Let w, be a walking angular frequency [rad/s] given by w, = 27r/T, All
variables with the subscript w are functions with respect to the angular
frequency of the input, w . A relative acute angle between the both legs at
the double support phase is called step angle which is notated by cyw, [rad].
If we can assume the linear relationship, w = LJ,, then the walking by the
leg length variation can be treated as a frequency response between the
leg length variation (the input) and the step angle (the output). Therefore
a walking gain W, is defined by the following equation such that a ratio
between the amplitude of the length variation:

4. Simulation
The control method can generate some walking patterns for the biped given
in section I1 through numerical simulations. The result for a cyclic motion
is illustrated in figure.2. The parameters of the biped are shown in table
1. m, and I, are mass and inertia, r,, sh, fm and sb identify the COG
position of each link as shown in figure 1. Note that the COG of the both
legs are forward, therefore, r f > 0 and r, = 0 in the table. Subscripts s, f
and b mean a shank, a thigh and a body, respectively. R and L are a radius
of feet and a default length of the legs, respectively. The amplitude and
the angular frequency for the leg length variation are K , = lO[mm] and
20 = 1.257r[rad/s]in this case. The average velocity converged to 0.4l[rad/s].

4.1. Walking gain

In the simulations, the radius of the feet is R = 0.78. It is the same length
as the initial length of the legs. Roughly speaking, the walking cycle, T,
is correspond to the cycle of the leg length variation for any w , because
the impact occurs when the both legs are the same as 0.78 [m] again at
t = n . T,/2. Therefore the phase property are not important. The gain
property is only analyzed as a frequency response. The walking gain and
the step angles are presented in figure 3. From the figure, a resonance point
is found. It appears when the velocity of the swing leg was correspond to the
 361

I I . I . I . I .I
-0.3 -0.2 4.1 O Indl
q

Fig. 2. A cyclic step of walking when K, = lO[mm], w = 1.25~[rad/s]. Left figure shows
angles and velocities of the stance and the swing leg. Right shows the phase diagram.

I I I
0.8 1 1.2 1.4 nlradl
Frequency of kncc bending w

Fig. 3. Walking gain and step angle

velocity of the stance leg. Therefore, the resonance frequency is smaller than
the characteristic frequency of the legs. Though the frequency shifts to left
with respect t o the amplitude K,, the gain characteristics with respect to
K , correspond except the region around the walking resonance frequency,
in the left of figure 3. Therefore the step angle can be regarded as a linear
function with respect to the amplitude of the leg length variation. (Remark:
since the feet length are not limited, the step angles around the resonance
frequency are too large.) There are appropriate frequencies which excite
biped walking for a desired step angle. Figure 4 shows energy consumption
with respect t o the input frequency. Figures 3 and 4 are compared, and
the higher the walking gain is, the more effective walking is in the sense
of consumed energy. Thus high gain frequencies are also appropriate for
effective walking.
362

I I . I . 1 . I - , . , . , . , I
0.8 1 1.2 1.4
Frequency of knee n [rad/sl

Fig. 4. Consumed energy

5 . Experiment
5.1. Experimental apparatus
We introduce our experimental apparatus called Prototype Biped Emu IV
(PBEmu-IV) shown in figure 1. A Direct Drive (DD) motor is equipped in
each hip joint so that it can rotate with few viscosity under no actuation.
Each knee joint is developed by a kind of prismatic joint which is composed
of a DC motor and a ball screw. Each foot is spherical whose center is
located at the hip joints. It is meant that the system is stable and it is easy
to keep standing upright position. In addition, the COG of the body is
located at the hip joint. The property makes PBEmu-IV walk more easily.
A parallel link is attached to the apparatus so that the lateral motion can
be constrained, the absolute position can be detected and the biped can
walk on a circle which radius is about 2[m]. An information of foot impact
is detected by a touch sensor embedded in each ankle which moves upside
down for 5[mm]. A sampling cycle for control is l[ms] and data are logged
each 100[ms]. The configuration for the apparatus are shown in Table 2.

Table 2. Configuration of PBEmu-IV

2.O[kg] l.O[kg] lS.l[kg]

radius 0.801m1 0.llml

 363

5.2. Experimental results of biped walking

0.1 l mk
6 E
E
Y
e<
8 0 0 8
2 -
Q

-0.1 -1z%!
0 10 0 Time[s] 10

Fig. 5 . Left figure: Relative angles of legs and absolute angle of the body. Right figure:
Angular velocities of the legs. Walking is stable in 5 steps (3.5[s]) and the average velocity
of the stance leg is converge to about 0.42[rad/s].

PBEmu-IV can walk using the leg length variation with the body control
when K , = 18[mm] and w = 1.8~[rad/s].The biped is standing upright
at the beginning, next the both knees are expanded 20[mm], and lastly
the legs start varying and the biped starts walking at the same time. The
experimental results are shown in figure 5. The figure presents angles and
angular velocity of the legs and the body. The origin is the position where
the biped stands upright. The walking gait is stabilized in five steps (3.5[s])
and the average velocity of the stance leg is converge to 0.42[rad/s]. The
body is kept around the upright position. The step angles have some vari-
ation which is caused by undulation of the floor. It should be remarked
that the convergence speed is much faster than the simulations. It is also
different from the analysis of McGeer that an eigenvalue of speed mode
approaches to unity with respect to feet radius increase.

5.3. Frequency response analysis f o r walking

Next, a frequency property of the biped walking is analyzed for the ex-
perimental results. As the result of the experiments, w, were constants
with respect to K,, and vary with respect to angular frequency of the knee
w. The relationship could be described w = w,. Therefore the previous
assumption is validated. The step angle a,, and the walking gain W, de-
noted by the equation ll are shown in figure 6. The figure is described in
and 6 5 K , I 18. a,, and W, increase when w de-
1 . 0 ~5 w 5 1 . 8 ~
creases and they are the largest when w = 1 . 0 ~Thus,
. the most effective
364

0.6 0.07 I , , , , , , , ,
0.06

-B
-a 0.4

I
f
m

0.2
’ ‘ I I ‘ ‘ I ’
1 1.2 1.4 1.6 1.B 0.01; 1.2 1.4 1.6 1.8
F q u e n c y of knee bending w n [radh] Frequency of knee bending w n [rad/s]

Fig. 6. step angle and gain.

frequency for our experimental apparatus is w = 1 . 0 in ~ the sense of the

walking gain. In the case, the velocity of the swing leg at the impact was
close to the stance leg. The large K, is, the larger a,, is. Graphs of W,
with respect to K , did not correspond each other, which is a little different
from the simulated results. We can also expect the existence of a walking
resonance frequency from the experimental results. However the step angle
around the resonance frequency is too large to walk according to the sim-
ulation. On the outside range of w and K,, indeed, the biped could not
walk. It is caused that the length of the feet are limited in the experiments.

6. Conclusion
For a planer biped, the frequency property was revealed for biped walking
by leg length variation. Some simulations gave the gain property and showed
a walking resonance frequency as a resonance point of the walking gain.
In addition, some experimental results were presented and validity of the
frequency analysis was verified. The frequency property showed appropriate
frequencies for biped walking in the sense of energy consumption.

References
1. T. McGeer, Passive Dynamic Walking, IJRR, Vol. 9, pp. 62-82, 1990
2. McGeer: “Stability and control of two-dimensional biped walking”, Technical
report CSS-IS TR 88-01. Simon Fraser University, Centre for Systems Science,
Burnaby, BC, CANADA, 1998.
3. S. Collins, et al. Eficient Bipedal Robots Based o n Passive-Dynamic Walkers,
Science, no. 5712, pp. 1082-1084, 2005.
4. M. Wisse et. al., “How to Keep From Falling Forward: Elementary Swing Leg
Action for Passive Dynamic Walkers”, I E E E Pansactions o n robotics, V01.21,
No.2, pp.394-401, 2005.
 Leg Control for Changing Locomotion between Leg-type and
Wheel-type Designs based on Effective Use of Total Power
Tokuji OKADAl, B. Wagner TANAKA2, Toshimi SHIMIZU’
Department of Biocybernetics, Niigata University
2Graduate School of Niigata University
Ikarashi 2-8050, Niigata-Shi, Japan 950-2181
IEmail: okada@eng.niigata-u.ac.jp
http://okada.eng.niigata-u.ac.jp

This paper proposes a method for determining joint control variables of a

mobile robot having the ability t o move using a leg-type or wheel-type design
when necessary. Specifically, we discuss the transient state of changing the
types based on minimization of total energy cost and distributing necessary
power into motors for actuating hip and knee joints so that the energy cost
of each type becomes equal. In both cases, we calculate the control variables
of the two joints uniquely by using an electrically calculated ratio of the hip
joint motor capacity to the knee joint motor capacity that are found in the
robot’s hardware design specification. Demonstrative simulation clarifies that
the type change is carried out continuously and smoothly without any need
for mechanical adjustment and manual assistance in a stationary state while it
stands in the same position and in a dynamic state while it moves in the same
direction of locomotion.

Keywords: Legged Locomotion; Wheeled Locomotion; Minimization of Power

Consumption; Energy Cost Function; Motor Capacity.

1. Introduction
Leg-type robots have high mobility on irregular terrain with adaptability
in climbing up and down stairs, and walking over obstacles. On the other
hand, the robot cannot move fast like a car even on a flat surface. The
legged robot proposed by Y. Zhou,l is a simple model of a rigid-link binary
walking robot that moves forward in a straight line with a wave gait. If
it has a hybrid mechanism with legs and wheels, the robot can move on
different terrain while steering itself. For example, the robot2 can select
either of wheel-type or leg-type electro-mechanically. Other robots having
legs and wheels a t its hinged leg ends3 can move on flooring and rugged
terrain. The wheelchair robot proposed by A. Gonzales et u Z , ~ is able to

365
366

Fig. 1. Definition of parameters and symbols. Configurations in (a), (b) and (c) are
seen in the movements of leg-type, change-round and wheel-type, respectively.

climb up a staircase.
This paper proposes two methods for changing the locomotion between
the leg-type and the wheel-type design without human assistance or me-
chanical operations using jacks, screwdrivers, monkey wrenches, levers and
so 0n5.6 The first is t o simply minimize the total power consumption. The
second is t o share the necessary power between the hip and knee joints.
In both cases, we take into consideration the ratio of the energy cost for
producing the hip joint torque to the cost for producing the knee joint
torque. Actually, this ratio is estimated using motor capacities found in the
robot’s hardware design specifications. For instance, such a robot driving
hip and knee joints by motor capacities of 3O[W] and 2O[W] makes a ratio
of 1.5. This value becomes greater when the motor at the hip joint is more
powerful than that at the knee joint.
We show that the joint control variables are determined uniquely by
applying the above methods and that the robot can change locomotion
type at the same position smoothly without bumping, swinging or sliding
on the ground. In the demonstrative simulation, we verify that the two
methods are advantageous in making the robot barrier free from obstacles
on the ground and in giving it high mobility in a real environment.

2. Definition of parameters for the motion analysis

Our PEOPLER is composed of 4 hip joints and 2 legs pivoted at each hip
joint rim with the same mechanism.6 Also, it is designed symmetrically on
the right and left side, and controlled with synchronism in motion of the
front and rear structures. Thus, one hip joint side view in Fig.1 is enough
to understand the robot’s motion. The left and right diagrams depict side
views of the leg-type and wheel-type design, respectively. The middle dia-
gram shows an instantaneous case belonging to the two types on a common
slope since it has two landing contacts. We call this case, change round
configuration. Parameters and symbols in the figure are helpful to analyze
 367

and extract equations for determining the joint control variables.

C: ratio of the Jh’S motor capacity to the J k ’ s motor capacity
c‘ : modified value of c for unusual motor use
d h : horizontal distance of the J k from the ground contact
d k : horizontal distance of the J h from the ground contact
h: height of the obstacle
1: leg length of the robot
r : radius of the knee joint rim
H : height of the hip joint, i.e. height of the robot body
J : joint ( J h ; hip joint, J k ; knee joint)
R : radius of a wheel
(a circular profile surrounding the two knee joints)
Th : torque of the hip joint
Tk : torque of the knee joint
W : axle load
7: leg posture angle (a pair of y have a bisymmetrical direction)
<: ratio of 1 to R
8: arm angle from the robot’s front direction
8, : inclination angle of the road surface
3. Leg motion analysis of the locomotion type change
In order to always maintain the robot in a standing position without distur-
bance from the scattered obstacles on the ground while it moves according
to the leg-type, we consider kinematics related to the joint control variables
of 8 and y with a torque of Th and Tk. When referring to Fig.1, we have
the following equations.

Th = Wdh = w ( r COS 8 - !I sin y) (1)

Tk = Wdk = Wlsiny (2)
+
H = r sin8 1 cosy, (3)
where in general H varies in the following range.

R<HIr+l (4)

3.1. P o w e r distribution f o r using input energy effectively

Motor capacity is an important specification in distributing electric power
effectively without heat loss. Evidently, the capacity shows maximal energy
consumption. Using more than the specified amount will shorten the mo-
tor’s life. Since power is consumed with the two motors, energy cost of a
368

d, d, dr dr Ed, 4d,

Fig. 2. Reasonable standing configuration when the load factors of the two motor o p
erations are proportional to motor capacities. di stands for a scale unit
small capacity motor is high when it operates with heavy loads. On the
other hand, the energy cost of a large capacity motor is low when it oper-
ates for the same load. Therefore, we define c as the ratio of motor capacity
of the hip joint to the knee joint for using power effectively.
This idea is clarified by the designs using three different values of c to
generate driving torque suitably for motor equipment. Let suppose that
c1, c:! and cg take the values of 0.5, 1.0 and 1.5, respectively. Also, suppose
that the total motor capacity, say 3O[W], is available to drive the two joints.
Then the configuration to keep the same hip joint height is illustrated in
Fig.2. The configurations in (a), (b) and (c) are generated under clrcp and
c3, respectively. In (a)1capacities of the hip and knee joint motors are 1O[W]
and 2O[W], respectively. That is, the knee joint can use much power than
the hip joint. In (b), the two motor powers are the same. In (c), they are
18[W] and 12[W] and the hip joint can use 1.5 times more power than the
knee joint.
3.2. Method f o r minimizing total energy cost
We define the energy cost of each joint for driving is reversibly proportional
to the motor capacity but proportional to the necessary torque that the
motor has to produce using reduction gear mechanisms. According to this
definition the joint using a large amount of motor capacity is less expensive
and the joint using a small amount of motor capacity is more expensive
for producing the same joint torque. Hence, we evaluate total energy cost
spent on the two joints by E,, then we have

+
E a = (Th):! ( c T ~ ) ~ (5)
To make the value of E, minimal it follows that
 369

Fig. 3. Solution to the robot becoming stuck using angular shift a.

Using Eqs.(l-2), one can get the 6th order equation of unknown parameter
sine. Solving the equation is not simple, but graphical tracing and explo-
ration gives the angle B through sine. Then y is calculated from Eq.(3).
Also, the six order equation produces the critical values of H and y
when 8 becomes zero. In fact,

One end of the two legs must be free from sliding on the ground. But
this approach is not practical and we thought about shifting 8 a little bit
from zero so that one of the two legs ends separates from the ground. This
shift, say a , prevents the legs from becoming immovable. Therefore, we
assign the shift a heuristically to 0 for determining y in the lower part.
y = cos-l(H - r s i n a ) / l (9)
Figure 3 shows the stuck robot configurations on the left and solved robot
configurations on the right by using the shift a(=n/s[rad]). A special
robotic sensor can sense an obstacle on the ground and surmise whether
it is possible to maneuver around the obstacle by checking the height h
defined by

h = qR, (10)

where q (0 < q < 1) means the coefficient of the self rolling up power of the
robot. Of course, the swing phase leg in the change round configuration is
also needed t o avoid the obstacle without pushing on the obstacle, however,
if the robot cannot avoid the obstacle it is better to step on it than go around
it. To make this possible the value of 0 in Fig.l(b) is assigned using the
height h to have

2r }
e=sin-ij R - ( 2 R - h ) - t a n 8 , ~ 4 1 2 ~ ~ ~ 2 0 , - ( 2 R - h ) 2
-e3 (11)

Also, the y is assigned by Eq.(3).

370

Fig. 4. Sequence of leg configurations based on minimizing total energy cost. Two
neighboring configurations differ in the hip joint height by 5[mm].
3.3. Method f o r bisecting motor power inversely t o the
energy cost
In such a case when the condition of the motor changes caused by such
things as overheating, excessive friction, frame weakness and etc., each mo-
tor's load factor might be changed. In general, it is meaningful t o distribute
enough power to each of the two motors so that Th = c' Tk is satisfied ir-
respective of the total energy cost. Unbalanced energy cost error, say Eb,
between the two joints is evaluated by

Similar to Eq.(6), the value of 8 making Eb minimal is calculated by

dEb/de = 0. However, it is difficult to solve the equation, since Th and
Tk change their signs, in general. Therefore, we simplified the equation as
follows by assuming that Th and Tk are positive. Therefore,

Insert Eqs.(l-2) into dEb/f% = 0 to obtain

r2(B-1)sin20-2BHrsinO+B(H2-12)+r2=0, (14)

where B = (c' + 1)2. Hence, we find

+
sine = { B H f d B 2 l 2 B ( H 2 - l 2 - r 2 )+ r 2 } / { r ( B- 1)) (15)

Introducing sine into Eq.(3) yields y. This method is useful t o assign the
modified ratio of c' when needed in unusual cases.
 371

Front

Fig. 5 . Overwritten leg motion in the period while the robot changes locomotion type.
Results of (a) and (b) are based on the total energy cost minimization method under
c=l and bisected power sharing method when H decreases with a small difference in
AH=3[mm], respectively.

4. Simulated motion results of the locomotion type change

We simulate robot motion over a transient period while the robot changes
locomotion from one leg type to the other type. Using overwritten configu-
rations and sequential snapshots we verify the motion based on the methods
explained in Section 3. Particularly, we demonstrate variation of the robot
configurations generated under the influences of c or c’

4.1. Minimization method of the total energy cost

Aforemensioned 6th order equation is utilized to determine the control vari-
ables to minimize total energy consumption at higher positions of the hip
joint from the ground. On the other hand, at lower positions of the hip
joint, the control variable 6 becomes a, and y is calculated from Eq.(9).
In Fig.4, the robot motion sequence is illustrated separately under c=0.5,
q=0.4, <=1.5 and a=~/12[rad].The sequence from (10) to (1) stands for
the locomotion change from the wheel-type to the leg-type design, rever-
sionary. In Fig.5 all of the transient configurations are overwritten. In the
left diagram, the robot’s center in the higher and lower positions is distin-
guished by the trajectories of G I and G2, respectively. In Fig.4, the robot
centers of GI and G2 appear in the configurations from (1) to (5) and from
(6) to ( l o ) , respectively. We notice that the mode is changed at the same
contact with the ground without any disturbance from the obstacle hav-
ing the height h. Therefore the robot makes a smooth locomotion change
without bumping or sliding. And one leg end does not slip on the ground.

4.2. Bisected power sharing method

In our power sharing control strategy, two motors’ load factors are assigned,
therefore power for driving the hip and knee joints are given independently.
372

In this case, the control variables of 0 and y are calculated using Eqs.(l5)
and (3), and the robot motion sequence is obtained. Fig.5b shows these
results under c’=l.O and <=1.5 by overwriting configurations. Evidently,
the configurations in the lower hip joint position cannot be determined
because there is no solution mathematically. The characteristics of H I R
versus 0 and y are obtained but not shown in the limited pages in number.
The left and right diagrams show results related to the methods of total
power saving and bisected power sharing, respectively.
5 . Conclusion
We proposed the two methods to use motor power effectively for driving
the hip and knee joints in the locomotion change between the leg-type and
wheel-type designs. In order t o evaluate energy cost for driving the two
joint motors, we defined the ratio between the two motor capacities and
showed two methods for calculating the two joint control variables using
not only the ratio but also the rate of leg length t o wheel radius. The first is
to minimize total energy cost effectively. The second is to bisect necessary
motor power so that each energy cost becomes equal. In unusual cases, the
ratio is modified gradually so that the energy cost is changed appropriately.
In the simulation of locomotion, we made it possible t o visualize the robot’s
motion on a display while it changes locomotion from the leg-type to wheel-
type design and vice versa. This simulation resulted in showing that the
locomotion change is carried out on a flat road without disturbance from
scattered obstacles. A demonstration using an experimental model and also
extension of the mode change on a slope are parts of our future work.
References
1. Y.Zhou: “On the planar stability of rigid-link binary walking robots”, in
Robotica, V01.21, no.3, pp.667-675, 2003.
2. S.Hirose and H.Takeuchi:“Studyon roller-walk”, Proc. IEEE Int. Conf. on
Robotics and Automation, pp.3265-3270, 1996.
3. K.Sonehara, T.Yamamoto and et.al.: “Development of Breadboard Model
for Locomotion of 3-Lewg Wheeled Type Robot”, Tech. Rep. of IHI, Vo1.44,
no.3, pp.241-247, 2004.
4. A.Gonzalez, R.Morales, V.Feliu and P.Pintado: “Improving the mechanical
design of a new staircase climing wheelchair”,Proc. of CLAWAR 2006, pp.14-
18, 2006.
5. T.Okada, W.T.Botelho, TShimizu: ‘Compatible use of a legged robot as a
wheeled robot and its demonstrative simulation”, Proc. of CLAWAR 2006,
pp.34-44, 2006.
6. T.Okada, W.T.Botelho, T.Shimizu:“Development of a rotating four-legged
robot, PEOPLER for walking on irregular terrain”, Proc. of CLAWAR 2003,
pp.593-600, 2003.
MOVEMENT SIMULATION FOR MERO MODULAR WALKING
ROBOT

ION ION', ION SIMIONESCU*, ADRIAN CURAJ3 ALEXANDRU MARIN4

'Technology ofManufacturing Department 'Mechanism and Robot Theory

Departn~ent,~Automatic
Control and System Engineering Department, 4Hidraulic
Department, POLITEHNICA University of Bucharest, Romania

E-mail: ioniSl@ yahoo.com

Modern methods of drawing up machines and tools necessarily include a simulation stage
of their functioning, which means to use a mathematical pattern of the real, original system.
These activate the functioning simulation that encompasses several rules and specifications
whose enactment generates behavior data and the instructions operating on the pattern's
description variables. The veracity and validity of a real system depend on the compliance
degree. The real system and the model are different by the fact that whereas for the former
the manner to generate conduct data can be completely unknown, for the latter they mean a
group of rules or specifications, whose enactment puts out the conduct data, namely
instructions operating of the model's description variables. In the paper, a robust model of
the direct and inverse analysis of the open kinematic chains of the walking robot legs is
constructed.

1. Introduction
Scientists of all times have been permanently mesmerized and have studied the
simplest but the most important movement, namely the mechanical movement of
humans and animals. Humankind is so much anthropomorphism addicted that it
is almost impossible for it to conceive or imagine automatic systems, even
having artificial intelligence, and that are not anthropomorphic.
Since long ago, many research teams worldwide have been focused on
goals such as creating an autonomous walking robot equipped with functions
like handling objects, locomotion, perceiving, navigation, learning, judgment,
information storage and intelligent control, and that can carry out tasks like
altering the multitude of the parts belonging to a dynamic universe.
The access of man to dangerous areas where his safety is jeopardized made
the scientific research approach topics of various purposes and conceive devices

373
374

that through their performances aim at covering different fields. The

architecture of these systems is quite different and depends on their purpose and
destination.
For example, walking robots protect the environment better because their
contact to the ground is discrete, which substantially diminishes the surface to
be crushed, the robot’s weight can be optimally distributed on the contact
surface through controlling the forces. The variation of the distance from the
ground allowed the robot to step over young trees or other vegetation growing
in the area it moves on.
The walk is defined by the manner the waking robot moves between two points,
under specific circumstances. To achieve and guide a walking robot requires
thorough knowledge about all walking possibilities because choosing the
number of legs and their structure depends very much on the selected walk. The
selection of the walk type depends on several elements such as:
shape and consistency of the ground the robot walks on,
walk stability driving and controlling the movements of the elements of the
walking systems,
speed and mobility movement requires.
It is very difficult to select the type of walk, mainly during real walking.
Therefore, it is necessary that the ground surface to be defined before selecting
the walk.
The walking robot’s steps are a sequel of movements of the legs, coordinated
with a succession of movements of the body for the purpose of moving the robot
from one place to another

2. The establishing the mathematical pattern of the movement system

Theoretical research of robot movement assumes nine important stages:
1. the establishing of the mathematical model of the movement of the
kinematic and dynamic system;
2. the process and construe the results obtained through simulation
with a view to determining the system’s conduct;
3. the structural and geometrical optimization of the movement
systems compounds;
4. the movement of compounds system;
5 . the establish the functional block scheme to calculate the
compounds;
6. the establishing of the test methodology to perform the system’s
functions that is need and the identity of the functions generators
walking requires;
 375

7. the design of the guidance system’s software;

8. the determination of the initial parameters and data characteristics
of the system’s structure and state;
9. the conduct of the simulation program.

Figure1 The Denavit -Hartenberg axis systcm attached of modular walking robot; it is suggested
support of technological equipments to work in farming and forestry.

2.1, Movement simulation by Denavit - Hartenberg formalism

Figure 2. The Denavit - Hartenberg coordinate axes systems for a leg mechanism
376

Let us a modular walking robot consist of three modules. [3]. Each module has
two 3-DOF legs, symmetrically arranged on the platform axis (fig.1).
The legs- on the right - onto the movement direction are superscript marked
with 2i, i = 1,6 whereas the legs on the left with 2i-1. Each platform of the rear
modules is connected to the platform of first module by a 3-DOF kinematic
chain with two links and three rotational pairs. The axes if these pairs are
concurrent and perpendicular two by two.
In order to carried out the movement simulation of a leg, a coordinate axes
system is attached to each link, with the Denavit - Hartenberg rule [2]. This
formalism may not only simplify the problem formulation, but can also yield
considerable advantage in the solution of simulation problem. The pairs of each
leg are numbered consecutively from A which is pair number 1 to C which is
pair number 3.
The Denavit - Hartenberg systems attached to each link are subscript
numbered as the pairs respectively. The platform is designed as link number (0)
and the remaining links are numbered consecutively. All pairs of the leg
mechanism are rotational and actuated ones.

2.1.1 Denavit - Hartenberg systems attached to the leg links of the walking
robot
The characteristic axis Z, of each pair should be defined. The positive sense of
each of these axes is defined arbitrarily. If the axes Z; and Z,.l are skew with
respect to each other, then there is one common perpendicular between them.
The perpendicular is designed as the 4, axis. If the Zi and Zi-, axes are parallel,
the 4. axis may chosen as any common perpendicular. The positive direction of
the X,axis is designed as proceeding from Zl-l to Zi. If the Zi.l and Z,.l intersect,
the positive sense ofX, axis is arbitrarily.
When the 4, axes are all defined, then are define both the Y, axes and the
origin of each right hand coordinate system. So, a coordinate system defined is
attached to each link. The parameter a; is defined as the distance from 0,Z; to
Oi+lZj+laxes, measured along O,+,X,+,.Because of the orientation of the
Oi+lXi+l.axis,a; is always positive.
The parameter ai is defined as the angle between the positive 02; and the
positive Oj+lZl+l axes, as seen from positive Oi+&+l.
The parameter Bi is the angle between positive O x j and the positive Oi+lXj+l
axes, as seen from positive 0;Z;.
 377

The parameter si is defined as the distance from O x i to Oi+lxi+laxes, measured

along the OiZi axis.
Under this definition, the Denavit - Hartenberg transformation matrix A has {
the well-known form:

0 0 0
cos e{ - cos a: sin e{ sin a{ sine: .
(1)
sin e{ cos a( cos e{ - sin a{ cos e j
0 sin a{ cos a:

To mould the walking robot’s moves is assumed that:

1. The kinematical length of the binary link (1) is null and it is connected
to the platform (0), by pair A and to the link (2) by pair B; the axis of
pairs A and B are perpendicular.
2. the binary link (2) is connected to the link (1) by the pair B and to link
(3) by the pair C, the axis of pairs B and C are parallel.
The A{ matrix performed the coordinate transformation of a point belonging to
link (3) from o{x:y4jz{ system to o:x{y(zJ’system attached to link (2). In
a similar manner, the coordinates of lower end point P belonging to link (3)
from O ~ x ~ y system
~ z { to o{x{qJz{system attached to the platform (0) is
performed by the equation

This matrix equation described the geometrical model of the leg 1 and 2 of
the walking robot. The goal of the direct kinematic analysis is to calculate the
position, velocity and
- acceleration of the end point P, in terms of the pair
variables ei, i = 1,3. In inverse kinematic analysis, matrix equation (2) is
solved with respect to the pair variablesei, i= c.
The positions of the point P and the positions of the platform with respect
to the reference coordinate axes system OXYZ fastened to the ground are
considered as known. Therefore, the position of the point P with respect to the
platform coordinate axes system are known.
378

The movement of the legs of the rear modules are controlled by the
following equations:

Each of these matrix equations is equivalent with three nonlinear equations and
has six unknowns, namely variables of the pairs. In the inverse kinematic
analysis three out of six unknowns must be imposed from independent
conditions.

Table 1.
 379

Fig. 3 shown the schemes of the sequence of the Denavit - Hartenberg

coordinate systems attached to the links of the left leg of the first module (a) and
to the links of the leg of the left rear module (b).

i 242
a

Figure 3 . The Dcnavit - Hartenberg axes systems attached to the links of the left leg of the first
module ( a )and of thc rear module (b)

In the movement simulation program, the parameters of the Denavit -

Hartenberg transformation matrices have the values in Tables 1 and 2.

2.1.2 First and second - time derivative of the pair variables

Through the repeated differentiated of the equation (2) with respect to the time,
yields:
380

(4)
In the inverse analysis, these equations are solved with respect to the first and
second - time derivative respectively of the pair variables.
The velocity and the acceleration components of the point P on the axes of
the coordinate system attached to the platform (0) are considered as known. In a
similar manner are differentiated the equations (3). -
ii
In these matrix equations, the variables e @ and ep, k = 1,6 of the pairs and
their derivatives with respect to the time are imposed by the walking simulation
program.
Because of these variables' particular values, the Denavit - Hartenberg
transformation matrices have the following simpler particular forms:

1 0 0 0
o cose:i -sine12i 0
A,2i =
o cOsefi o
s; 0 0 1

1 0 0 0
2i
cos eii cos O i i - sin 0,
2i 0
A, = sin O i i sin eii cos 2i e2 0
4 0 0 1
 381

(5)

Keeping stable is a special problem that occurs while the robot walks, when one
or more legs are in the transfer phase. When all the legs are in the support phase,
it is obvious that the protection of the center of gravity is within the support
polygon.
If one or more legs are in the transfer phase, the geometry of the support
polygon changes and it occurs the risk that the protection of the center of gravity
moves outside the support polygon.
Solutions to such situations depend on how the modular walking robot is
configured. In fig. 4 are shown some sequence of the computer simulation of the
gait of the modular walking robot

Figure. 4 Computer graphics simulation the gaits of the modular walking robot

Conclusions
The movement simulation of the walking robots may be idealized into a
mathematical model for the purpose of kinematic analysis. The techniques of
idealization can play the decisive role in easiness, precision and time of calculus
for the problem solving. The Denavit - Hartenberg method is numerically
382

robust, the solutions are either exact in the sense that is possible to refine them
up to an arbitrary accuracy. A modular walking robot could have one or more
modules. The motions of the legs must be coordinated so that the conditions of
the gait stability of the system to be ensured.

References
1 A.P. Bessonov, N.V. Umnov, The Analysis of Gaits in six-legged Robots
according to their Static Stability, Proc. Symp. Theory and Practice of
Robots and Manipulators, Udine, Italy, 1973
2. Denavit J., Hartenberg R.S., A Kinematic Notation for Lower Pair
Mechanisms Based on Matrices, Journal of Applied Mechanics, Tr. ASME,
1955, Vol. 77
3. I.Ion, I. Simionescu,. A. Curaj, MERO Modular Walking Robot Support of
Technological equipments, The 8thInternational Conference on Climbing
and Walking Robots, September 12-14,2005 London, UK
4. I.Ion, I. Simionescu,. A. Curaj, Mobil Mechatronic System With
Applications in Agriculture and Sylviculture.The 2th IFAC International
Conference on Mechatronic Systems ,December 8-12,2002-Berkeley -
USA pp94 1-946.Pergamon Press
5. R.B. McGhee, G.I. Iswandi, Adaptive Locomotion of a Multi-legged Robot
Over Unarranged, IEEE Trans. On Systems, Man, and Cybernetics, Vol.
SMC-9, No. 4, April 1979, pp. 176-182
6. S.M. Song, and J.K. Waldron, An Analytical Approach for Gait Study and
its Application on Wave Gait, International Journal of Robotics Research,
Vol. 6, NO. 2, 1987, pp. 60-71
7. J.K. Waldron, Modeling and Simulation of Human and Walking Robots
Locomation ,Advanced Scholl Udine ,Italy 1996
 Trajectory Generator for Rhythmic Motion Control of Robot
using Neural Oscillators

Weiwei Huang, Chee-Meng Chew, Geok-Soon Hong and Nithya Gnanassegarane

Mechanical Engineer-ing, National University of Singapore, Singapore

This paper presents a structure of coupled neural oscillator which can be used as
central pattern generator(CPG), for humanoid robot locomotion. Parameters
tuning is always a big challenge for oscillators. In this article, some properties
of neural oscillator, like limit cycle requirement, amplitude and frequency are
studied. The objective is to develop a trajectory generator which has simple
parameters’ tuning. This architecture is suitable in various fields where rhyth-
mic motion is required. We tested the oscillator structure on a simulated biped
robot. The result demonstrates a stable walking behavior in both forward and
backward straight walking and different walking speeds.

Keywords: Neural Oscillator; CPG; Limit Cycle; Trajectory Generator; Biped

Locomotion.

1. Introduction
Exploiting the nature of human walking is always an open question for
human locomotion researchers. Neurophysiological studies of animal loco-
motion have revealed that the basic rhythm of movement is controlled by
rhythm generating networks in the nervous system, which are called cen-
tral pattern generator(CPG).l A CPG is a network of neuron oscillators
coupled by mutual inhibition which can be used to control the rhythmic
movement of a walking robot. The goal of the CPG research is to generate
the desired motion for the joints. Many interesting works based on neural
oscillator have been done to explain such automatic oscillatory activities.2-6
In these research works, the oscillators’ parameter tuning is the key chal-
lenge, especially for complex CPG network with multiple neurons. Some
researchers have addressed this problem by tuning the parameters math-
emati~ally.~ Another approach is to use optimization tools like GA8v9 to
search for the desired parameters. Other researchers focus on deriving the
learning algorithms in order to adjust nonlinear oscillators automatically.10
However, tuning the oscillator parameters is still a problem in the CPG

383
384

research.
In this paper, we present the analysis which enable us to understand and
tune the oscillator with ease. The details are organized as follows: In Section
2, limit cycle behavior of neural oscillator is discussed and mathematical
formulas for oscillator frequency and amplitude are given. In Section 3,
the oscillator structure is designed to generate the desired trajectory. In
Section 4, we empirically verify the robustness of the proposed structure by
numerical simulation.

2. Limit Cycle Behavior

In this section, we will study the limit cycle behavior of the neural oscil-
lator model proposed by M a t ~ u o k a A
. ~ new approach based on PoincarB
Bendixson theorem and piecewise linear analysis is proposed. We introduce
the approximate formula of amplitude and frequency to help tuning the
oscillator parameters.

2.1. Stability Analysis

The model of the two neurons Matsuoka's oscillator is:3

71?i1 = c - u1 - pv] - u [ u 2 ] + - c hj[Sj]+ (1)

72i)l = ['U1]+ - 211 (2)

TIC2 = c - u 2 - pv2 - u[.1]+ +c hj[gj]- (3)

T27j.2 = [.2]+ - w2 (4)

[u]+
= max(0, u) [u]-
= min(0, u ) (5)
Y = [.1]+ - [u4+ (6)
where u 1 ( 2 ) represents the inner state of the neuron; q ( 2 ) is the degree of
neural adaptation; c is the constant stimuli; 71(2) is the time constants; p
is the parameter that indicates the effect of adaptation; u represents the
strength of inhibition connection between neurons; g j is the feedback from
environment and Y is the oscillator output.
PoincarBBendixson Theorem is one of the important results for limit
cycle behavior analysis of nonlinear dynamics. In particular, it can be used
to establish the existence of limit cycle in the state space.
Poincarb-Bendixson Theorem:ll Suppose that
(1) R is a closed, bounded subset of the phase plane;
 385

(2) x = f(z)is a continuously differentiable vector field on an open set

containing R;
(3) R does not contain any fixed points; (Note:Fixed point is a point that
satisfies f (2.) = 0);
(4) There exists a trajectory C that is confined in R; (i.e. a trajectory that
starts in R will stay in R for all future time)

Then either C is a closed orbit or it spirals toward a closed orbit as t --+ 00.
That is, R contains a limit cycle.
Based on the theorem, we will prove the existence of limit cycle in neural
oscillators step by step. For better understanding, we begin our proof with
condition (2) and condition (1) will be proved later. Here, we assume that
there is no sensor feedback to the oscillator. That is , g j = 0.
Satisfying of condition 2: The oscillator model is not continuously dif-
ferentiable, because [u]+ = rnaz(u,0) cannot be continuously differentiable.
In order to meet the smoothness requirement, we used a smoothed version
of the function by first replacing the rnaz(u,O)function with the func-
tion w,
f(u) =5
and then by replacing the absolute value with the function
arctan(ku). When k 00, the vector fields of f (u)and I u
I
are the same. Since the function f ( u ) is continuously differentiable, [u]+
can be considered as continuously differentiable.
Satisfying of condition 3: It is difficult to check whether there are fixed
points or not in equations (1)-(6). To overcome this, we separate the u1,
u2 region into four subset quadrants ((u1 1 0,212 2 0 } , (u1 2 0,uz <

0}, (211 < 0, uz 2 0}, (u1 < 0,uz < 0)). For each respective quadrant, the
equations (1)-(4) can be transformed into a linear equation:

T172Y + (71 + 72 - U 7 2 ) i . + (1+ p - U ) Y = 0 (7)

T1TzY + + T 2 ) Y + (1 -k p)Y - C = 0
(71 (8)
T1T2Y + + + (1 + p)Y + C = 0
(71 72)Y (9)
Y=O (10)
where Y = +]I.[ - [uz]+
Equation (7) has a fixed point at u* = (0,O). It can be proved that
when 1 +2 - a < 0, equation (7) has a negative damping and the fixed
point is a repeller. Therefore we can modify the region with the punctured
region R as shown in Fig.1 which does not include (0,O) point."
Equation (8) has a fixed point a t u* = (+, c(1 - $)). The require-
ment for u* t o be a fixed point is 1 - & < 0, because it has to be within
the quadrants (u1 2 0,uz < 0). If 1 - & > 0 that is u < 1 p, then u* +
386

is not in the region. This implies that u* cannot be a fixed point. We can
also give a similar proof for Equation (9).
Here we have proved that when a - 1 - > 0 and a - 1 - ,6 < 0 are
satisfied, then the output Y has no fixed point. If the output has no fixed
point, this implies that the state variables also have no fixed point. On the
other hand, if the state variables have k e d points, the output will converge
and maintain a constant value.
Satisfying of condition 4: To ensure that a confined trajectory ex-
ists, we construct a trapping region R, which is a closed connected set
such that the vector field points inward everywhere on the boundary
of R as shown in Fig.1. Then all trajectories in R are said to be con-
fined. Matsuoka has proved that the oscillator state is b ~ u n d e d , ~ t h aist
~ 1 ( 2 ) ~ ~ Z L 1 ( 2 ) ( 0 ) ) + C + ~ J Z 1 1 ( 2 ) ( 0 ) ) where I~l(Z)(O)l
and Ivl(2)(O)l are initial condi-
tion. We construct a larger boundary for R in order to show the vector
fields clearly, where bound = c Pvmaz au,,,, + + Vmax and Umax are the
largest value of q ( 2 ) and ul(z).

Fig. 1. Trapping region R

Satisfying of condition 1: The state region R shown in Fig.1 is a bounded

subset of the plane.
Therefore the neural oscillator model described by equations (1)-(6) has a
unique limit cycle in the bounded region R if
71
a-1-->0
72
 387

The trajectory is unique in region R based on the Existence and Uniqueness

Theorem. l1

2.2. Frequency and Amplitude

Frequency and amplitude are two important properties of a periodic signal.
We found that when la-1 -2 I is small, two neurons are always in the active
state, that is in the region (u1 2 0,212 2 0). Generally, we can approximate
the oscillator frequency F by using the damped natural frequency of the
linear system in equation (7) as follows:

2c
UAmp =
1+P+a
There are several advantages in deriving these expressions. First: with
the knowledge of both frequency and amplitude, we can know the shape of
the oscillator output. Second: we can set the parameters of oscillator based
on the desired frequency and amplitude. Third: from the formula we found
that if we fixed the value of a , b and p1the value of 7 2 is directly proportional
to the frequency and c is directly proportional to the amplitude. Therefore
changing the performance of the oscillator becomes easy.

3. Trajectory Generation
In this section we will present the neural oscillator structure that we have
used to generate periodic trajectory. First we present the structure of the
pattern generator which is made of coupled neural oscillators and then we
discuss those properties of the structure that make it suitable for rhythmic
motion control.

3.1. Structure of the Trajectory Generator

In the previous section, we found that both amplitude and frequency could
be modified by changing certain parameters. When la - 1 - 21
is small,
the shape of the oscillator is similar to the sine wave. According to Fourier
series, an arbitrary periodic function can be decomposed into a weighted
sum of sinusoidal components. It is plausible to have the idea of using
several oscillators to generate any desired trajectory.
388

Fig. 2. Network structure of the oscillators

Fig.:! shows the structure of this network. We associate to each oscillator

a phase adjustment for the difference in phase of each oscillator with respect
to the first oscillator in the network, thus enabling us to correct any phase
error between the oscillators. The equations describing this structure are
as follow

TilCil = C - Uil - - U [ U i 2 ] + - [pi]' -c hij[gij]+ (15)

Ti27jZl = [.ill+ - 2r21 (16)

TilCiz = C - U i z - pWi2 - U [ U i i ] + + [pi]- +c hij[gij]- (17)

Note: p i is the difference in phase of ith oscillator with respect to the first
oscillator.

3.2. Property Analysis

In this section, we present several numerical experiments to verify the prop-
erty of the neural oscillator structure. Firstly, we set the structure to gen-
erate a knee trajectory which is obtained from the normal human walking
experiment. After Fourier analysis, we found that the trajectory is mainly
composed of three different frequencies components. Therefore we assign
three oscillators t o generate this trajectory. Fig.S(a) shows the compari-
son between the reference and the generated knee joint trajectory. If we
want to reduce the number of oscillators to simplify the structure, we can
choose the main frequency components. Because of the intrinsic properties
 389

of stability, the structure can produce trajectories that are stable t o per-
turbations(Fig.3(b)). When we consider the perturbation as a feedback to
the oscillators, then the oscillators tend to adjust the output automatically.

(a) (1),(2)The reference knee joint trajectory and (b) (1)Stability against perturbation given
its FFT; (3)generated knee trajectory through g i j , (2)performance of each oscillator

Fig. 3. Result of trajectory generation and stability against perturbation

Another important aspect of this neural oscillator structure is that it allows

easy modulation of the amplitude and the frequency of the trajectory. Since
the frequency and amplitude are linearly related to 72 and c, simple modu-
lation of these values can cause a large variation of stable trajectories. Some
properties are shown in Fig.4. Because of the properties of coupled oscilla-
tors, modulation of these parameters is always smooth and thus interesting
for robot trajectory generation.

4. Empirical Verification
In this section, we will test the structure on a biped robot in dynamic
simulation environment. We assign the oscillators to the task space of a
3D robot, which has a significant reduction in the number of parameters
compared to joint space. It was demonstrated that robust steady walking
can be achieved with a proper arrangement of oscillator^.^
As shown in Fig.5, this structure is composed of three oscillators which
control X,Y,Z direction movement. The three oscillators are coupled to-
gether by the phase adjustment. Y direction oscillator generates the refer-
390

.. . ... .. ..... . . . . . .. . . . 2

I
50 im
,
lio 1UI.
I
150
0 50

(a) Modulation of amplitude (b) Modulation of frequency

Fig. 4. Smooth response of the parameter change

ence phase signal to adjust the phase of X and Z direction oscillators. The
proposed structure has multiple oscillators. However, we choose a simple
case when there is only one oscillator to control each direction. In the sim-
ulation, roll and pitch angle of the robot are used as feedback to adjust the
output of the structure.

(a) Proposed control architecture using (b) NUSBIP-I1

neural oscillators

Fig. 5. Simulation model

The simulation is carried out in Yobotics environment which is a Java

package for simulating fully dynamic systems. The structure of the robot
model is the scale down model of ~USBIP-II(Fig.5(b)).The body height
is 36cm; the foot length is 8cm and the whole weight is 5.84kg. In Fig.6
both forward and backward walking are demonstrated. Backward walking
was achieved by changing the sign of the X direction oscillator. The Os-
cillator output and sensor feedbacks are given in Fig.’l(a). In Fig.i’(b), the
 391

phase diagram of the Y direction motion are plotted. The trajectory starts
from the origin and eventually converges to a constant oscillatory swinging
motion.

Fig. 6. Snapshots of forward walking (Step length=0.02m, Velocity=O.O8m/s) and back-

ward walking (Step length=0.012m, Velocity=O.O48m/s)

The simulation shows a robust walking behavior. The forward walking

speed is 0.08m/s and backward speed is 0.048m/s. Since the size of the robot
is small, the walking speed is acceptable compared to the previous work.
In the simulation, 'u' direction oscillator is able to adjust the phase of the
other two oscillators. This implies that our structure also shows the ability
of entrainment property. This is our preliminary work and for simplicity,
we have generated trajectories with only one frequency component. In the
future, we plan to demonstrate a more complex oscillator structure and the
robustness of the structure will be studied.

-1s - O P l ' ' ' ' ' ' * I

I 0 5 < 16 * 25
-mi
3 -004 -003 4 0 2 -001 0 001 002 009 004
Y Dimdim Position

(a) q z , qv, qz entrainment and roll, pitch feed- (b) The limit cycle behavior of frontal plane
back in the straight walk

Fig. 7. Simulation results

392

5. Conclusion
In conclusion, this paper presents a new way of analyzing t h e neural oscil-
lator. A simple trajectory generator structure is designed. We analyze the
properties of this structure through numerical and dynamic simulations.
The robot shows a robust walking behavior with three coupled neural oscil-
lators. In t h e future, we may consider implementing the structure to achieve
more complex walking behaviors. We will also explore more efficient feed-
back pathway t o achieve robust walking.

References
1. S. Grillner, Neurobiological bases of rhythmic motor acts in vertebrates, Sci-
ence,vol. 228 1985, pp. 143-149.
2. K. Feng, C.M. Chew, G.S. Hong, T. Zielinska, Bipedal locomotion control
using a four-compartmental central pattern generator, Procs. of 2005 IEEE
International Conference on Mechatronics and Automation,Vol. 3, pp. 1515-
1520.
3. K. Matsuoka, Sustained oscillations generated by mutually inhibiting neurons
with adaptation, it Biological Cybernetics, vol. 5 2 , 1985, pp. 345-353.
4. G. Taga, A model of the neuro-musculo-skeletal system for human locomotion
I. emergence of basic gait, Biological Cybernetics, vol. 73,1995, pp. 97-111.
5. G.E. J Nakanishi, J. Morimoto and G. Cheng, Experimental Studies of a
Neural Oscillator for Biped Locomotion with QRIO, Procs. of 2005 IEEE
International Conference on Robotics and Automation, pp.596-602.
6. M. Williamson, Neural control of rhythmic arm movements, Neural Networks,
~01.11,1998, pp.1379-1394.
7. A. Arsenio, Neural oscillator networks for rhythmic control of animats, From
Animals to Animats MIT-Press, 2000.
8. D. Garis, Evolution of a time dependent neural network module which teaches
a pair of stick legs to walk, Procs. of 1990 9th European Conference on Arti-
ficial Intelligence, pp. 204-6.
9. M.A. Lewis, A.H. Fagg, AS Solidum, Genetic programming approach to the
construction of a neural network for control of a walking robot, Procs. of
1992 IEEE International Conference on Robotics And Automation, vo1.3, pp
2618-23.
10. L. Righetti and AJ Ijspeert, Programmable Central Pattern Generators: an
application to biped locomotion control, Procs. of 2006 IEEE International
Conference on Robotics and Automation, pp.1585-90.
11. S.H. Strogatz, Nonlinear Dynamics and Chaos with application to physics
Biology, Chemistry, and Engineering, (Addison-Wesley Publishing Company),
2000.
12. W. Huang, C.M. Chew, G.S. Hong, Neural Oscillator Analysis for Rhyth-
mic Control of Robot Locomotion, Procs. of 2007 ZEEE/RSJ International
Conference on Intelligent Robots and Systems, submitted.
Observer-based control of a walking biped
robot: stability analysis

V. Lebastard, Y. Aoustin, and F. Plestan

IRCCyN, Ecole Centrale de Nantes, CNRS, Universitk de Nantes - Nantes - France
{Vincent.Lebastard, Yannick.Aoustin,
Franck.Plestan}Qirccyn. ec-nantes .fr

Abstract
The considered problem in this paper is the estimation the absolute orienta-
tion of a dynamical stable five-link biped during a walking gait. The solution
is founded on a nonlinear observer based on the super twisting approach, such
that the estimation error converges in finite time and only uses the measure-
ments of the joint variables. This observer is coupled to a nonlinear control
law, the stability of the cyclic walking gait under the observer-based controller
being also proved.

1 Introduction
The interest t o take into account the imbalance phases in the walking gait of
a biped is real in order t o make its locomotion more fluid and faster. However,
such improvements induce the use of the whole state of the biped in the de-
sign of reference trajectories and the control law computations. For example,
accurate sensors are necessary to measure the joint variables, the absolute
orientation of the biped in imbalance phases. Unfortunately, a precise mea-
surement of the absolute orientation of a walking robot in these imbalance
phases is, by a technical point-of-view, quite difficult to get. For the joint
variable measurement, usually a gearbox reducer leads to a good accuracy.
But to estimate the orientation of a walking biped in imbalance phases with
for example accelerometers, gyrometers, inertials units many others problems
have to be taken into account in the measurement results such as the offset er-
rors, the impacts effects, the temperature effects, the bandwidth limits of the
sensor versus the dynamics of the walking biped, the robustness, the price ...
For these latter reasons, there is a real interest to explore the possibilities of
observers in order t o estimate absolute angular positions and velocities from
only the knowledge of the relative angular variables.
Very few works have been done for the design of such observers, the previous
works on observers design being done especially for the estimation of velocities

393
394

(for noiseless differentiation) by supposing that all the angular variables are
measured [8]. First results on the design of observer/controller using only the
measurement of joint link angular variables (relative angles) for three-links
and five-links biped robots with no actuator in the ankles, in both cases of
its stabilization in a vertical position and its walking, have been proposed by
the authors and are based on Kalman filter [l],high gain observers, (high
order) sliding mode observers [6, 51. This latter class allows t o establish the
finite-time convergence of the estimation error and, through the high order of
sliding mode, removes the chattering effect while preserving robustness and
improving the accuracy.
The paper proposes a new version of second order sliding mode observer,
based on super twisting approach [7], in order t o estimate the absolute orien-
tation of a five-link biped without feet during a walking gait. The imbalance
phases have been taken into account when the biped is in single support. This
algorithm is using only measurements of the joint variables, and not yet its
time derivatives as previous works, which implies less noise in the controller.
Furthermore, this paper presents the proof of stability of observer-based con-
trol, which has been still done only for a 3 links-biped robot: this result is an
extension of [6].

2 Model of a planar five-link biped robot

The considered biped is walking on a rigid and horizontal surface. The robot
is modeled as a planar biped. It consists of a torso, hips, and two legs with
knees but no actuated ankles (see Figure 1). The walking cycle takes place
in the sagittal plane and consists of successive of single support phases and
impacts. The complete model of the biped robot consists of two parts: the
differential equations describing the dynamics of the robot during the swing
phase, and an impulse model of the contact event (the impact between the
swing leg and the ground is modeled as a contact between two rigid bodies).
The dynamic model of the biped in single support phase between successive
impacts is derived from the Lagrange formalism

D(q,el)4' + H ( q , 919 + G(q) = Br (1)

with = [q3i q4i q32 942 qiIT1 G e l := [q3i q32 q4i q42lT, the vector joint angles
r
vector, and = [Ti r2 r4IT
r3 (see Figure 1 ) l , D the inertia matrix, H the
centrifugal and Coriolis effects matrix, G the gravity matrix. The torques u
are applied between the torso and the stance leg, the torso and the swing leg,
at the knee of the stance leg, and a t the knee of the swing leg, respectively.
Then, the model can be written in state space form by defining [3]

x= [ 9
D-'(-HQ - G + B r ) 1
Leg 1 is the stance one, leg 2 the swing one.
 395

with x = [qT qTIT. The state space is taken such that x E X c RIO = {x =
[qT q. T]T I q E N , q E M } , where N = (4 E R5 I 14.1 < q~ < m} and
M = ( - T , T ) ~ .The walking surface is taken as x E S = {x E X I z 2 ( q ) =
0 , &(q) < 0 ) with zg(q) the altitude of the swing leg tip and its time derivative
& ( q ) . An impact occurs when the swing leg touches the walking surface,
also called the ground. The impact between the swing leg and the ground is
modeled as a contact between two rigid bodies. Under the same hypotheses
and results of [4], the final result gives an expression for x+ := (q,df) (the
state value just after the impact) in terms of x- := ( q , 4 - ) (the state value
just before the impact), which is expressed as x+ = A ( x - ) . The overall biped
robot model can be expressed as a nonlinear system with impulse effects [3]

x = f(x) + g ( q r e L ) r , for x- 6S , xf = A ( x - ) , for x- E S. (3)

3 Controller and observer design

3.1 Controller design

The control for the walking gait [4] consists in tracking the four joints reference
angles q41d, q32d and q42d of the biped. During the single-support phase,
the degree of under-actuation equals one, as only four outputs can be driven.
Then, the robot gets a walking motion if the controller drives to zero the
output vector h(x) defined as (with 0 defined in Figure 1)

h ( x ) := [q31 - q31d(0) q32 - q32d(0) q41 - q41d(0) 442 - q42d(o)lT

The control strategy consists in decoupling the system and in forcing the
system to evolve by arbitrarily stated dynamics through the use of a finite
time convergent control law [6].

3.2 Analysis of observability

Consider the dynamical system (2), with y the vector composed of the mea-
sured variables y := [y1 yg y 3 y4IT = qreL

+
x = f(x) g(y)T, y= cx (4)
In the biped context, this model describes the swing motion and is studied
over one step, i.e. for t E [Ti,T;+'[, with Tj (resp. Tjfl) the initial (resp.
final) impact time of the step i. As g(y)T, the input-output injection term of
(4), is fully known, an observer for (4) can be designed by the following way.
With abuse of notation, consider the next nonlinear system, which is the part
of (4) without the input-output injection term g ( y ) r ,

x = f(x), y = cx (5)
The observer displayed in Subsection 3.3 is designed from this model (5).
396

Proposition 1 ([S]). There exist observability indices [kgl k32 kll k42IT and
7 c X such that system (5) is observable for x E 7 .
From Proposition 1, system ( 5 ) is locally equivalent t o the canonical form,

i = Az + cp(z), y = C,z (6)

with ( A , C , ) Brunovsky form, under state transformation z = @(z) =
[yl . . . y$-) . . . y4 . . . yy4-l)]T .

3.3 Step-by-step observer

In this section, the observer is based on triangular form one [2] and its main
property is the finite time convergence to zero of the estimation error. System
(6) is still on (particular) triangular form. A such observer of (6) can be
written as [2]
+
f = A i ~ ( 2 ) E(t)X(.)+ (7)
with f E 2 the estimated vector of z . Knowing that the principle of this
class of observers consists in forcing, each in turn, estimated state variables
to corresponding real ones, in finite time, this latter property is based on an
adequate choice of E ( t ) and x(.),i.e. E ( t ) and x are defined such that the
estimation error e = f - z converges to zero in finite time. The originality in
the following is the use of the super twisting algorithm [7] in order to ensure
the finite time algorithm: this allows to not use time derivative of estimation
error in the computation of the observer.

Finite time convergence observer

In a sake of clarity, and without loss of generality, only the observer design for
a third order system is fully displayed in the sequel, 2.e. z E 2 E R3. Then,
in this case, system (6) reads as

Zl = zz, i z = 2 3 , i3 = f 3 ( z ) , y = z1 (8)
with z = [z1 z2 ~ 3 1 Then,
~ . an observer for (8) reads as

21 =fz + El(t)Xl(.),j 2 = 23 + E2(t)XZ(.),.& = f 3 @ ) + E3(t)X3(.) (9)

with Ei(t) and xi (1 5 i 5 3) defined such that each estimation error ei =
f i - zi converges to zero in finite time. To ensure a finite time convergence,
the function xi(.) is based on the super twisting algorithm [7] and reads as

xi = Q + Is~I'/' sign(Si), <i = ai sign(Si) (10)

with

Zi - f i for i > 1
 397

Determination of Ei and sketch of proof of finite time convergence

From (8)-(9), the dynamics of estimation error, ei = i i - zi,reads as

) , e3 + Ez(t)x2(.),63 = f 3 ( 4
4 = e2 + E I ( ~ ) x ~ 6(2. = - f3(z) + ~%(t)x3(.).
0 Step 1. Suppose that el(0) # 0, and observer (9) is initialized such that
El = 1 and Ez = E3 = 0. The error dynamics reads as

dl = e2 +XI, e2 = e3, e3 = f 3 ( i ) - f3(z) (12)

As x1 is based on the super twisting algorithm with appropriate gain value,
el reaches zero in finite time at t = tl. Then, V t 2 t l , e l ( t ) = e l ( t ) = 0 ,
i.e.
el = 0, el = e2 +
x1 = i~ - z2 x1 = 0 + (13)
From (13), one gets 2 2 + x1 = z2 and z2 = 2 2 .
Steps 2 and 3. The proof takes the same way than step 1.

Loss of observability and observation algorithm

As mentioned in [6], during the swing phase, along the nominal trajectories,
for each observability indices possibilities, there is loss of observability. This
problem has been solved through the use of two observers structures based on
different observability indices [6].

3.4 Proof of stability

In this section, the objective is to prove the stability of the walking of the
biped controlled by previously presented controller with finite-time observer.
The stability can be proven on the basis of a restriction of the PoincarC’s
map to an one-dimensional manifold [4]. In [6],an original extension of [4] to
observer-based controlled systems has been proposed for a three-links biped
robot. In the sequel, an extension is made for five-links biped robots. AS the
“real” state vector is not fully measured, the L‘rea”’zeros dynamics and impact
surface manifolds can not be used in the stability proof, the idea is then to
suppose that the estimated state is on “estimated” zeros dynamics and impact
surface manifolds. The finite-time convergence of the observer and controller,
at the end of the first step, ensures that the estimated manifolds are the same
than “real” ones. Then, it is possible to use the standard reduced PoincarPs
approach to establish, over the second step, the stability.

Notations. The impact time at the end of the ith step is noted Tj and is
+
taken as the time origin for the (i l ) t h step. Let TA denote the convergence
(towards 0) time of the estimation error over the ith step and T&the conver-
gence time of the controller over the ith step. The observer and the controller
have been tuned such that TA 5 T& < T j , i.e. the observer converges faster
than the controller over the step i. As the observer converges to real state in
398

finite time, it is obvious that TA = 0 for i > 1: then, from TA,one has P = x.
Numerical procedure. Let 2 denote the “estimated” zero dynamics mani-
fold, 2 = {P E X I h(P) = 0 , h(f) = 0}, S the “estimated” impact surface
manifold, S = {P E k I z 2 ( P ) = 0, &(P) < 0}, A? := {P := [$,4’]’}, 4 and
q being the estimated values of q and 4. The conditions required to define the
restricted Poincard map are
1. S n 2is a smooth sub-manifold of A?. It is equivalent to the fact that the
T
map [ / ~ ( i ) 22(P)]
~ has constant rank equal to 9 on S n 2,which
is obvious to prove. If (i,
2 ) E S n 2,2 equals a constant, denoted PO.
Let y := [h(P)T&?)IT which has full rank at PO (for definition of 0, see
Figure 1). On 8, one has h(P) = 0

Proposition 2. Let us define p , a difleomorphism from R + 3 n 2 to

complete equation (14) with the configuration vector ijo of the impact such
as
p ( Q )= [ [yp]
40

At each configuration Go is associated an unique angle 8 0 .

2. The cross section for the Poincard maD will be taken to be S . the “es-
timated” impact surface. Define X : R*+ R computed by the’following
manner.
Let 8: = 8+(T:) (resp. 80) denote the initial estimated angular veloc-
ity (resp. position) just after the initial impact. Compute P+(T:) =
[cj:i:IT := p ( 8 $ ) , the estimated state vector of the biped after the
impact. State the real state after the impact as (given that q41,
q32r q42 and the corresponding velocities are measured)

Z+(T,O) := [q&(80) q i l ( e 0 ) d 2 ( 8 0 ) 4 i 2 ( 8 0 ) q1(T,O)

431 (T;) 441 (TT) 432 (T;) 442 (Tf) 41 (T;)]

P+(m := [ddio, dl(80) 4%9o) q&(eo) Gl(~0)

T
Q;l(eO 7 8;) 4il ( 6 0 , Q,., 4;2 ( 8 0 , 422 ( 8 0 , Q,., 41 (T,O)]

Use x+(Tf) as the initial condition in (3) controlled by a control law

which uses 2 . Simulate until one of the following happens :
 ’ 399

a) There exists a time T: for which 1 2 = 0, .& < 0 and TA 5 TA <

T j < 00 (recall that, for t 2 TA, ?(t)= z ( t ) ) :apply again the im-
pact model to ?-(T:). Note that, fotThen, ?+(T:) = A(?-(T:)).
At this time, the real and estimated state variables have same val-
ues, viewed that the observer has finite-time convergence of esti-
mation error, and that the observer gains have been tuned such
that TA < TI’. Use ?+(T:) as the initial condition in (3). If there
exists a time TI”for which 1 2 = 2 2 = 0, .& = i 2 < 0 such that T 2 <
TZ < TI”< 00, then apply the impact effect ?+(T;) = A(?-(T,)). 9-
One gets X [e+(T:)]:= &(TI”)= p-l [?+(TI”)]; else X [e+(T:)]is
undefined at this point.
b) There does not exist a TI” > 0 such that 1 2 = 0, & < 0; in this
“
case, it is also true that X B^+(T:)]is undefined at this point.

4 Simulations

The parameters of the five-links biped prototype “Rabbit” [3] are used to
design the five-links biped parameters. The masses and lengths of the links
(Indices 31, 41, 32, 42, 1: swing leg (femur, tibia), stance leg (femur, tibia),
torso, resp.) are m31 = m32 = 3.2 kg, m41 = m42 = 6.8 kg, m l = 17.0528 kg,
131 = 132 = 141 = 142 = 0.4m, 11 = 0.625m. The distances between the joint
and the mass center of each link are ~ 3 =
1 ~ 3 =2 0.127m1~ 4 = 1 ~ 4 =2 0.163m,
s1 = 0.1434m. The inertia moments around the mass center of each link are
131 = 132 = 0.0484kg.m2, 1 4 1 = 142 = 0.0693kg.m2, I1 = 1.8694kg.m’.
The inertia of the rotor for each DC motor is I = 3.32 lop4 ko.m2. The ratio Y

N of the gearbox reducers equals 50. Value U of the applied torques equals
150 N . m .
In order to determine if the closed-loop system is stable, function X is
evaluated for B^,$ E [0.5,4], q1 E [7r/12 - ~ / 3 0rad,7r/l2 7r/30 rad] and +
q1 E [0.02 rad/s, 0.22 rad/s]. Figure 2 displays function X : X is undefined for
8:(T:) less than 0.825 radls and more than 3.75 radls. A fixed point appears
at approximately @(T;) = 1.665 rad/s, and corresponds to an asymptotically
stable walking cycle. Figure 3 displays e+(T,“)with respect to the initial “real”
kinetic energy E, = Ec(q31, q 3 2 , q41,442,41,q31,$32,q41rq42,q1) and its initial
“estimated” value 8, = E c ( 6 3 1 , 6 3 2 , 6 4 1 , 6 4 2 , 6 1 , a 3 1 , 6 3 2 , 6 4 1 1 6 4 2 , 4 1 )andthe ini-
tial estimation errors on unmeasured variables q1 and 41. This representation
allows to know the possible initial estimation errors on all the variables which
are admissible in order to guarantee the convergence to the limit cycle through
the kinetic energy. For each point of this 3D-area, one ensures that each esti-
mated state variable converges to the corresponding one, the control outputs
reach zero before the end of each step, and the biped robot converges to a
stable limit cycle. In terms of this 3D-area, the final stable walking cycle cor-
responds to the red star point. Figure 4 displays the walking motion of the
400

biped robot as a series of stick figures over three steps on the stable limit
cycle.

5 Conclusion
The first contribution of this paper is the design of an original finite time
observer, based on second order sliding mode approach, which estimates the
absolute orientation of a dynamical stable five -ink biped. This task is im-
portant and vital in order to control its locomotion with high accuracy. The
stability of the nonlinear association control/observer is proved for cyclic walk-
ing gait, which is a hard task and an originality in the litterature. The next
step will consist in experimentally testing the strategy on prototype Rabbit
[3]. An other extension of this work consists in estimating the orientation
of humanoid robot in 3D when its feet sole is partially in contact with the
ground, or when the ankles torques are dramatically limited.

References
1. Y. Aoustin, G. Garcia, and A. Janot. Estimation of the absolute orientation of
a two-link biped using discrete observers. In Proc. Mechatronics and Robotics
Conf. MECHROB, pages 1315-1320, Aachen, Germany, 2004.
2. T. Boukhobza and J.P. Barbot. High order sliding modes observer. In Proc.
IEEE Conf. on Decision and Control CDC, pages 1912-1917, Tampa, Florida,
USA, 1998.
3. C. Chevallereau, G. Ahba, Y. Aoustin, F. Plestan, E.R. Westervelt, C. Canudas
de Wit, and J.W. Grizzle. Rabbit: a testbed for advanced control theory. IEEE
Control Systems Magazine, 23(5):57-79, 2003.
4. J.W. Grizzle, G. Abha, and F. Plestan. Asymptotically stable walking for biped
robots : analysis via systems with impulse effects. IEEE Transactions on Auto-
matic Control, 46(1):51-64, 2001.
5. V. Lebastard, Y. Aoustin, and F. Plestan. Absolute orientation estimation for
observer-based control of a five-link walking biped robot. Robot Motion and Con-
trol: Recent Developments, Lecture Notes in Control and Information Sciences,
335, 2006.
6. V. Lebastard, Y. Aoustin, and F. Plestan. Observer-based control of a walking
biped robot without orientation measurement. Robotaca, 24(3):385-400, 2006.
7. A. Levant. Sliding order and sliding accuracy in sliding mode control. Znt. J. of
Control, 58(6):1247-1263, 1993.
8. P. Micheau, M.A. ROUX,and P. Bourassa. Self-tuned trajectory control of a biped
walking robot. In Proc. Int. Conf. on Climbing and Walking Robot CLAWAR,
pages 527-534, Catania, Italy, 2003.
 401

c
*.

Fig. 3. @(T,")versus initial kinetic en-

ergy estimation error E and initial esti-
mation error on q1. All the initial con-
ditions in the 3D-box allow tha conver-
gence to the stable limit cycle character-
ized by the red star.

Fig. 1. Schematic of biped robot: abso-

lute and relative angles - Cartesian coor-
dinates.

Fig. 4. Plot of walking as a sequence of

stick figures.

Fig. 2. Function X (solid line) and iden-

tity function (dotted line) versus @ ( T j ) .
This graph establishes the existence of an
asymptotically stable walking motion.
 OPTIMIZATION OF HUMANOID ROBOT MOTION DURING
THE ELEVATION OF AN OBJECT

HAMED AJABI NAEINI

Biomedical Department, Azad University Sciences and Researches Branch, Tehran
,Iran

MOSTAFA ROSTAMI
Biomedical Department, Azad University Sciences and Researches Branch, Tehran
,Iran

This research is aimed at generation optimal motions of two dimensional humanoid robot
while robot must elevates an object to hypothetical height with knee bending and without
knee bending from the ground. The approach is demonstrated by using a 4 and 5 link
planar biped. The optimization problem is dealt with using Pontryagin's Maximum
Principle. As the biped is essentially submitted to gravity forces, the motion is generated
by minimizing the joint actuating torques and time. After implementation optimization
technique, a final two point boundary value problem is solved by applying a shooting
method. The approach presented is illustrated by various numerical simulations applying
to a four and five segment planar robot which has four or five active joints during
elevation.

1. INTRODUCTION
In the field of biped locomotion,our reference model is human gait.Although
the kinematics organization of human motion is quite easy to describe, the
underlying dynamics is complex to analyze and even more difficult to simulate.
If we want to move a robot, a simple means to define its motion consists in
introducing a kinematic model of elevation. It would not ensure the quality of its
dynamics. Synchronization with restrained values of joint actuating torques, as
well as moderate energy consumption is needed.
The prevailing idea is based on the minimization performance criterion with
a dynamic content involving essentially the joint actuating torques, or the energy
consumption.
To deal with the optimization problem, which is stated within the frame of
the optimal control theory. Accordingly, the implemented optimization technique
is the Pontryogin Maximum Principle (PMP). Kinematic constraints are taken
into account by means of an augmented integral criterion, and are dealt with

402
 403

using a penalty meth0d.h this way, the constrained optimization problem is

transformed into a sequence of unconstrained problems which are solved as two-
point boundry value problems.

2. KINEMATICS AND DYNAMICS OF A PLANAR BIPED

The planar multibody system more free joint are considered, and all of them
are powered. A complete Hamiltonian dynamic model is formulated on the basis
of a set of relative joint coordinates.

2.1. Kinematic model

We describe a sagittal model of an anthropomorphic humanod. This model
is made up of 4 and 5 limbs. In accordance with the schematic representation in
figure 1. we define the dimension and comulementarv notations :

Beginning of M Q t i m End of Motion Beginning o f Metion End o f Motion

4-DOFRDBDT 5-DDF ROB01

Figure 1. 4&5-link planar humanoid at first and end of elevation

q = (q,,....., q,)',vector of j o i n t coordinates

q = (q,,....., q , )',vector of j o i n t velocities
q = (q,,....., q, )',vector of j o i n t accelerations

2.2. Dynamic model

As we intend to use the Pontryogin Maximum Principle (or PMP)for solving
the dynamic optimization problem stated in what follows, let us recall that the
implementation of the PMP requires theb formulation of the dynamic model in
state space form. As indicated in reference [l, 21 the Hamiltonian dynamic
model not only best fulfills this requirement but, as well, strengthens the
404

robustness of algorithms used to solve the optimization problem. We present the

outlines of the formulation we need.
Firstly, introducing the Langrangian of the mechanical system
1 .T
77974) = 2 9 4 9 1 9 (1)
Where V stands for the gravity potential, and T is the kinetic energy defined
as

M being the (nx n) mass matrix of the kinematic chain, Lagrange's

equations of motion may be derived as

+ Q ; ~, i = I ,...,n (3)
Secondly, defining the conjugate momenta

dL
P,=- i = 1, ..., n (4)
84;
and the Hamiltonian

H(p,q) = PTil- L(q,il) (5)

Lagrange's equations in (4) may be reformulated in Hamiltonian form

iq.' =aH
3Pi

Or

With this formulation, Hamiltonian equations are ideally structured for applying
the Pontryogin Maximum Principle.
The double set of vectorial equations (7) can be recast as the 2n-order
differential vector equation

X(t) = F ( x ( t ) )+ Bu(t) = F(x(t),u(t)) (8)

 405

Where F is a nonlinear function in X, and B is the constant (2nx n) matrix .

Let us mention that a complementary transformation remains to be achieved
in order to perfect formulation (8). In consist in rescaling all the variables of the
problem to homogenize their order of magnitude. To that end, one can introduce
of the following reference quantities:L,M,T,I,Q, respectively length, mass,
time, moment of inertia and torque of reference which can be define and linked
as
- 1 1
M = - ( m , + ...+m,)
n
J=z 2
= -(QP"""" + ... + Q:.max
n
)

3. RESULTS

Simulated motion of robot from initial to end of elevation and relative joint
coordinates and angular velocities and anglular acceleration of joints during
lifting of anobject in following figures being represented :

Figure 2.0ptimal motion with and without knee flextion

Figure 3.Evolution ofjoint angle

406

Figure 4. Evolution of joint angular velocities

-*I
.......Od
- -6l
(b(

lb6
-0I'
- -c7
...... 8%)

.-.-br

bnrl8l linie#

Figure 5. Evolution ofjoint actuating torques

Used method in this research made a lot of way in solving. Modeling and
optimal control problems , and its result can be used in other biomechanical
researches.
According to figure 3 , the maximum range of joint angle is seemed in
second joint (pelvis) and third (shoulder) .
According to figure 4 angular velocity of joints in first and end of motion is
equal zero and shapaf ang ular velocity can consider a sine wave. According to
figure 5 , the maximum torque is created in first joint (ankle) and in start of
motion , (90Nm). This torque is comparable with created torque in pelvis joint at
bending position (48Nm). Also angle of first and third joints in first of motion is
more than its quantity in end of motion.
 407

References
1. M. Rostami ,G. Bessonet,”Sagittalgait of a biped robot during the single
support phase.”,Robotica,200 1
2. B.Colbert and F.Multon, “Biomechanical Simulation of Human Lifting” ,
Bennes University, France, 2001
3. L.L.Menegaldo, A.D.T.Fleury, H.I.Weber, “Biomechanical Modeling and
Optimal Control of Human Posture”, Journal of biomechanics, 36: 1701-
1712,2003.
4. W.Park, B.J.Martin, S.Choe, D.B Chaffin, M.P.Reed, “Representing and
Identifying Alternative Movement Techniques For Goal-Directed Manual
Tasks”, Journal of biomechanics, 38:5 19-527,2005.
5. M.Parnianpour and etc., “A Computational Method For Simulation of Trunk
Motion”, Biomechanics Research Laboratory,Yale University School of
Medicine, Jun 2000
6. William H.Press, Saul A. Teukolsky, William T.Vetterling & Brian
P.Flannery, “Numerical Recipes in FORTRAN: the Art of Scientific
Computing”, Press Syndicate of the University of Cambridge, 2th edition,
1986
7. H. Asada and E. Slotine , “Robot Analysis and Control”,NewYork ,John
Wiley,1986
Postural stability control for Robot-Human cooperation for
sit-to-stand assistance.

V. Pasqui*, L. Saintbauzel and P. Bidaud

Universite' Pierre et Marie Curie-Paris6, F R E 2507, ISIR
18 Route du Panorama 92 265 Fontenay-aux-Roses
*E-mail: pasquiOrobot.jussieu.fr
http://lrp6.robot.jussieu. fr/

J. Graefenstein
Leibniz University of Hanover
E-mail: j.gmefenstein@web.de

This article presents a fuzzy controller, for a robotic device, to ensure stability
of the user during the assisted sit-to-stand transfer. The first problem to be
addressed is the postural analysis of the chair rising. Experiments, with healthy
subjects, were performed with this aim in view. Analysis of external forces
shows that sit-to-stand transfer can be subdivided into several phases. The
observation of the Center of Pressure and of the horizontal component of the
handle force yields rules to observe the stability of the patient and consequently
adjust the robotic interface motion to the human volontary movement. These
rules are used in the fuzzy control implementation. The controller is validated
on experiments with healthy subjects and diseaded patients.

Keywords: Assistive device; robotic interface; human centered robotic; postural

stability; sit-to-stand; fuzzy control.

1. Introduction
The aim of the work presented in this paper is to realise a robotic interface for
equilibrium assistance during Sit-To-Stand (STS) transfer. Here, it is supposed
that interactive robotic devices, as human-centered robotics, is more comfort-
able and more efficient that traditionnal technical devices [l].
Robotics technologies have been investigated in the last few years to prevent
falls by a postural control of patients and to promote safe mobility [2], [3], [4],
[5], [6], [7], [8], [9]. But these robotic devices have no postural correction to
restore equilibrium.
Based on an analysis of the most common walking troubles associated with
aging or cerebellar syndrome we have designed and developed a robotic device
(in Figure 1) [lo] to help the patient to sit-down, stand-up and walk [ll].The
assistive device handles guide the patient to rise from a chair or to sit down,

408
 409

following trajectories which are based on parameters reflecting personal strate-

gies [12].

Fig. 1. Robotic Interface prototype

We propose here to detail how an adapted control can give interactive ability
to this robotic interface.

2. Method
By interactivity, we intend the capacity to interpret the postural movements
detected by the sensors t o trigger the movement or to maintain the postural
equilibrium.
To observe the postural state, experimental dynamical analysis of the stand-up
.
have been done in our laboratory (Figure 2) Results show different phases of
chair rising, that are matching with physiological 1itteratu:e [13]. Each phase
depends on interaction forces between human and handle : Fh = (Fho, Fhy); or
human and ground : Fg = (Fsa,Fga,) and their time variations. Reaction force
between human and ground is computed at Center of Pressure (COP) which
position may be used as a stability criteria [14].
The observation of the COPposition and direction of the force $h yields simple
rules to identify unstability cases or desired movement to trigger ( i s . beginning
of the STS). Fuzzy controller is used in intention detection to control neural
prostheses [I51 or orthosis using FES [16].
Fuzzy controller seems to be a good way for interactivity, then we have extended
the role of the fuzzy controller from the detection of voluntary movement to
the detection of the unstability.
The fuzzy control has to fulfill two tasks, thats defined two output:
e output 1: recognition of the current phase,
e output 2 : determination of proper reaction to ensure stability of subject.
 ___~ -_

Fig. 2. Different sit-to-stand phases analysis

The following fuzzy sets were defined for the output 1 (in Figure 3) : seated,
returned, preaccelaration (preac), acceleration (wc), start rising (startr), rise.
The _detection of the phases of th2 STS is obtained analysing the value of the
Fh, F, and the time variation of Fh.
"Returned" identifies the c a e when subject aborts stand-up and returns to
the seated position.

Fig. 3. Membership functions for output 1

 411

The membership functions for the output2, that determines the movement, if
proper phase is detected, are shown in Figure 4.The following fuzzy sets were
defined:
instable (inst) : object underlies high unbalance. Quick reaction is required.
stabilize (stab): object indicates desire of stabilization.
no move (nm): no movement is necessary in the horizontal direction.
adjust (adj) : object desires another position of the handles.

Fig. 4. Membership functions for output 2

If we denote H for high, Z for zero, L for low, EL for extremly low and EH for
extremly high, we have for exemple :

IF F,,=EL A N D Fh,=L A N D %=H THEN the human is RISING.

IF Fh, =H A N D CoP=L THEN the human posture is stable.

The rulebase for the fuzzy control is presented in the table figure 5 .

Fig. 5 . Rulebase for the fuzzy control

412

Detection of unstable posture is illustrated in figure 6, where both patient and

robot are modelised by a 3 links model each. The difference between these two
models is in the interaction with ground. We assume that robotic interface
cannot loose contact with ground while patient could if he is unstable.

Patient Robot i?

Fig. 6. Interaction between patient and robotic interface

If a subject, under perturbations, is verge on to loose his balance, he quickly

shifts the load within the foot support area in the opposite direction of fall
direction. If the impending fall directed forwards, the COP will rapidly move
in the same direction. An according reaction could be observed for a fall back-
wards.

3. Application
The complete structure of the controller is shown in figure 7.
The preprocessing block receives forces measurement, applies a filter and calcu-
lates the position of COP and its time derivatives. These outputs are processed
by the fuzzy logic block to identify patient posture state. Then, the correspond-
ing control mode is selected between those:
N o r m a l : tracking trajectory,
I m p e d a n c e : Impedance control according to the efforts of interactions mea-
sured,
Stabilization : modification of the tracking trajectory to stabilize the patient,
R e t u r n : the interface returns to the initial position,
The movement is triggered by the preacceleration phase.
For cases identified as a patient aborting movement, the robot returns to the
initial position. If postural unstability is detected the device motion in vertical
direction is stopped and a new desired position is computed that guarantee
patient stability.

4. R e s u l t s
The presented prototype is currently in Bellan Hospital for a rehabilitation
protocol validation. Many diseaded patients, with cerebellar syndrome, have
tested the device. This kind of pathology imposes a ballast walking-aid to filter
 413

P ntral

Y
to the robot

Fig. 7. Control structure

Fig. 8. Prototype in Bellan Hospital

shaking. The robotic interface is more comfortable, less tiresome and easy to
drive, In addition to assist in position change and walking, this device can
detect onset of fall. Then the robotic interface will response by changing handle
414

position, producing a force to balance the patient, as it is shown in Figure 9.

Fig. 9. Unstable posture corrected by the robotic interface

5 . Conclusion
Fuzzy logic is very usefull for detection of movement intention and unstable pos-
tures. Provided with a fuzzy supervisor, the robotic interface becomes highly
interactive.

Fig. 10. A commercial product

 415

Now, objectives are to develop rehabilitation protocols with collaboration with

the medical team of Bellan Hospital. A second objective is to produce a com-
mercial product with ROBOSOFT society. Such a product would be designed
as in Figure 10.

References
1. R. reiner, L. Lunenburger and G. Colombo, human-centered robotics applied
to gait training and assement., in Journal of rehabilitation research and de-
velopment, (5)2006.
2. S. C. Lee, K. Oh K. and J. Lee, A system for gait rehabilitation : Mobile
manipulator approach., in Proc. IEEE Int. Conference on Robotics and Au-
tomation, (ICRA’O2), (Washington, USA, 2002).
3. K. Nagai, I. Nakanishi and H. Hanafusa, Assistance of self-transfer of patients
using a power-assisting device., in Proc. IEEE Int. Conference on Robotics
and Automation, (Taipei, Taiwan, 2003).
4. Korean eldercare robots are coming. (2006), http: //robotgossip.
blogspot.com/2006/07/korean-eldercare-robots-are-com%ing.html.
5. M. Spenko, H.-Y. Yu and S. Dubowsky, A robotic personal aid for the mobil-
ity and health monitoring for the elderly., in IEEE Trans on Neural Systems
and Rehabilitation Engr, 2007.
6. N. Rea, G. Lacey, C. Lambe and R. Dahyot, Multimodal periodicity analysis
for illicit content detection in videos., in The 3rd European Conference on
Visual Media Production (CVMP 2006), 2006.
7. B. Graf, M. Hans and D. S. Rolf, Care-o-bot ii development of a next gen-
eration robotic home assistant., in Autonomous Robots, 2004.
8. J. Pineau, M. Montemerlo, M. Pollack, N. Roy and S. Thrun, Towards robotic
assistants in nursing homes : Challenges and results., in Robotics and Au-
tonomous Systems, (3-4)
9. H. Park, H. Hong, H. Kwon and M. Chung, A nursing robot system for the
elderly and the disabled., in International Journal of Human-friendly Welfare
Robotic Systems, (4)
10. P. MBdCric, V. Pasqui, F. Plumet and P. Bidaud, Sit to stand transfer assist-
ing by an intelligent walking-aid, in Proc. 7th International Conference on
Climbing and Walking Robots, (CLAWAR ’Od), (Madrid, Espagne, 2004).
11. P. MCdCric, V. Pasqui, F. Plumet and P. Bidaud, Elderly people sit to
stand transfer experimental analysis, in Proc. 8th International Conference
on Climbing and Walking Robots, (CLAWAR’OS), (London, UK, 2005).
12. V. Pasqui and P. Bidaud, Bio-mimetic trajectory generation for guided arm
movement during assisted sit-to-stand transfer, in Proc. 9th International
Conference on Climbing and Walking Robots, (CLAWAR’06), (Bruxel, Bel-
gium, 2006).
13. R. Aissaoui and J. Dansereau, Biomechanical analysis and modelling of sit
to stand task: a literature review, in Proceedings of IEEE International Con-
ference on Systems, Man, and Cybernetics, 1999.
14. P. Sardain and G. Bessonet, Forces acting on a biped robot. center of
416

pressure-zero moment point, in IEEE Transactions on Systems, Man and

Cybernetics, (5)2004.
15. P. C. Sweeney, G. M. Lyons and P. H. Veltink, Finite state control of func-
tional electrical stimulation for the rehabilitation of gait, in Medical and
Biological Engineering and Computing, (2)2006.
16. S. Hussein and G. M.H., Intention detection using a neuro-fuzzy emg classi-
fier, in IEEE Engineering in medecine and biology, 2002.
 RESEARCH ON UNDERACTUATED DYNAMICAL WALKING
OF 3D BIPED ROBOT

SHENG TAO, MA HONG-XU

College of Mechatronic Engineering and Automation, National University of
Defense Technolou, Changsha, Hunan Province, China
A new underactuated 3D biped robot and its control strategy are presented. Its hybrid
dynamic model is developed and gait is planned using method of time-invariant. By the
feedback controller and finite-time stable control strategy, the robot realizes stable
dynamical walking. Simulation results show that a stable limit cycle of dynamical
walking is achieved, and proposed control strategy is feasible. Based on analyzing the
relationship between walking velocity and robot configuration, a velocity control strategy
is presented and validated by simulation.

1. Introduction
Biped robot is always a research focus because it has the same appearance as
people and can do things instead of people. After many years' development, this
domain has got a lot of achievements. There are many famous projects such as
Honda P series biped robot, HRP series of Japan Trade and Industry (METI) I,
SDR series and QRIO biped robot of Sony. Leg Laboratory of MIT is one of the
pioneers in this area, Flamingo, Turkey, M2 were developed since 1980s.
Despite the nearly half-century's efforts, no biped robots have made their
way into sectors where their utility exceeds their novelty. It is conjectured that
one main factor contributing to so slow development of usable biped robot is the
low energy efficiency. It costs ten times or more energy than people when
walking the same length. In order to improve the energy efficiency, the research
in passive walking appeared and originated by McGeer in the late 1980s. In his
seminal work [2], McGeer built a four-link planar passive walker and performed
a detailed parameter variation and stability analysis. With his success, plenty of
researches on passive walking were carried out [3], [4].
Passive wallung is stable in theoretic, but in practice, vast work should be
done to adjust its parameters before realizing stable walking. Furthermore, the
robot should walk on a slope, so it has little actual significance. Based on the
passive walking, Grizzle begins to study underactuated planar biped robot. The
robot has no sole, contacts with ground by point feet. It maintains dynamic
balance by changing the contact states of the two feet. Grizzle in [5] proved that

417
418

the planar biped robot has a hybrid zero dynamics and gave theorems to estimate
the stability. But the lateral stability was not dealing with in his work. Rabbit is a
prototype of five links planar biped robot [ 6 ] ,which is the most successful one at
present. Underactuated planar biped robot can only walk around an origin
because it needs a no-actuated rotating bar to keep its lateral balance. Catherine
E in [7] and J. Maxwell in [8] proved respectively that the lateral stability of
human need active control. Steven H [9] and D. Kuo [lo] respectively designed
their 3D passive walking robot to perform 3D passive dynamic walking. But up
to now, there are little reports about 3D underactuated biped robot.
Based on the work of Grizzle, we design a 3D biped robot. The robot uses
feet with lateral distributing sole to realize lateral moment control. So the lateral
motion is controllable completely and the sagittal is underactuated. It can realize
3D dynamic walking. The remainder of the paper is structured as follows:
Section I1 develops the dynamical model, which is not continuous and includes
single support phase and double support phase. Section I11 develops the time-
invariant strategy to plan the gait of the robot. Section IV introduces the
feedback controller and uses finite-time stable control strategy to control the
walking. Simulations are done in Section V to validate the performance of the
control strategy and the stability of the robot. At last, a strategy is introduced in
section VI to control the walking velocity.

2. Dynamical Model
The robot we consider is shown in Fig. 1 It consists of 7 links: two identical
femurs, tibiae, feet and a torso. Each knee joint is a revolute joint with one DOF
(Degree Of Freedom). Each hip consists of revolute joint with 2 DOF. The sole
is a cylindrical bar, distributing in lateral plane. Ankle lies in the center of the
sole, and provides lateral moment with the robot.

\
Figure 1. Underactuated 3D biped Robot
 419

We make several assumptions to the robot as in [ 5 ] , [ l 11:

1. Walking takes place on a flat surface;
2. The walking is symmetric in the steady state;
3. The double support is instantaneous and involves the impact of the swing
leg with the ground;
4. The impacts are plastic and there is no rebound or slippage;
5 . The impulsive forces may result in an instantaneous change in the velocities,
but there is no instantaneous change in the configuration;
6 . In the single support phase, the stance leg acts as a pivot in sagittal plane.

2.1. Single Phase Dynamical Model

A biped robot in the single support phase is an open chain manipulator and made
up of tandem joints. A set of generalized coordinates of the robot is
q = [q,. . . q 9 ] ' , q, is the sagittal DOF of the ankle, which is an absolute angle
between the robot and the ground. q2 ...q9 are the relative coordinates of the
robot: q4 and q5 are the sagittal and lateral joints of the standing hip; q6
and q, is the lateral and sagittal joints of the swing hip; q9 is the lateral joint of
the swing foot.
Letting x := [qT,qT]' , applying the Euler-Lagrange's equations, the
dynamical equations of the robot in single support phase can be written as:

D(q) is the inertia matrix and always positive definite.C(q,q) is the Coriolis
matrix and G(q) is the gravity terms. The matrix B is a constant matrix used to
indicate whether a joint is actuated. Our robot has 8 active joints, and q, is
unactuated, then B = [O,,, I,IT . u = [u, ...us]' are the applied torque at the joints.

2.2. Double Phase Dynamical Model

The double support phase is the end of the current step and the start of next step.
As the assumptions 3-5, the double support phase is assumed to be
instantaneous and consists of two distinct processes: impact and relabeling. In
former a plastic impact takes place between the swing foot and the ground, and
the later is responsible for changing the roles of the two legs so that they can
share the same dynamical model.
The result of the impact and relabeling of the states can be expressed as:
X+ = A(x-) (2)
420

where “-”indicates the state before the impact and “+” after it. The detailed
process can refer the literature [ 11, 141.

2.3. The Complete Dynamical Model

The complete dynamical model of biped robot is:

x = f ( x ) + g(x)u xes
(3)
X’ =A(x-) X€S

where S is the switching surface, by the assumptions, it is the level ground.

3. Gait Planning
By far, the most popular algorithms are time dependent in planning the gait of
biped robot. The central idea of those algorithms is: planning the trajectories
q(t)of each joint, and tracking them by controller to realize walking control. But
for underactuated biped robot, we can’t control the time of walking directly and
the synchronization of time is not realizable, so the robot can’t walk stably under
those algorithms.
Time-invariant algorithm is employed in this paper to plan the gait of biped
robot. Time-invariant algorithm based on the definition of the reference
trajectory for M outputs (where M is the number of actuators), not as a function
of time, but as a function of a configuration variable 8 . 8 is monotone and
independent of the M outputs. We call those outputs as virtual constraints, as in
[15]. To a dynamic system of N DOF, defining M virtual constraints, it can be
reduced to a N - M dimension zero dynamics system. We can analyze the
stability of the whole system by studying the zero dynamics system.
During walking, lots of variable can be selected as 8 . In this paper, we
select the angle between the hip and the standing ankle in sagittal plane, as in
Fig.2.When the femur and the tibiae have the same length:

Figure 2. The left one is the pose looking from side, and the right is from back.
 42 1

e=q,+q,/2 (4)
When walking, one observes that the torso is maintained at a nearly fixed
angle, the hips remain roughly centered between the feet, and the end of the
swing leg traces an approximately parabolic trajectory6.The hip sways left and
right slightly and the two legs are always parallel. All of these characters can be
expressed by some parameters, such as the angle between the torso and the
uprightness ( Q:Iio), the length of standing leg ( L, ), the maximum of B ( OM" ),
the height of the swing foot ( h ) and the swing angle of the hip ( Q;:, ). As
mentioned before, the biped robot has 8 actuators, and 8 independent outputs are
need to restrict the walking. We define them as yI ...y8 . They are:
1. Keeping the torso upright in the lateral plane;
2. Of:, is a parabola trajectory;
3. The swing leg parallels the standing leg in lateral plane;
4. The swing foot is level in lateral plane;
5 . Q;:o keeps a fixed angle;
6. h is parabola trajectory;
7. L, keeps a fixed length.
8. The center of the hip is in the middle if the two feet in sagittal plane.
where 1-4 are the lateral constraints and 5-8 are the sagittal constraints. They
can express the walking pose of the robot basically. All of these can be
expressed as the output:
y=bl ... y8r= q e )
The dynamical model of biped robot is Eq.(l) and Eq.(2), its output is
(5)

Eq.(5). Then the task of the controller is to settle the output in finite time so that
the robot can impact the ground with desired configuration. All of those are
geometrical constraints to the configuration, not to velocities. So the walking
velocity is not control directly, it would differ from the desired one. The strategy
to control the walking velocity will be studied in the later section.

4. Control and Stability

4.1. Feedback Controller Design

The period of a step is finite, so controller should settle the system in finite time.
The control objective is driving the outputs to zero. Since the outputs only
depend on the generalized positions, and the dynamic model is second order, the
relative degree of each output component is at least two. Using standard Lie
derivative notation direct calculation yields:
422

d 2 y /dz2 = L>H(q,q ) + L,L,H(q)u

LfL,H(q) is decoupling matrix and always reversible. When the input is:

u(x) = (LgLfh(x))-'(v
- L>h(x)) (7)

The nonlinear system can be linearized as a double integral system. The

feedback functions used here come from [ 131:

with pi = yi + (1/2 - a)sign(s ji)I&. jiI2-" , 0 < a < 1 E > 0 are the
parameters to control the settling time

4.2. Stability of the Controller

Westervelt and Grzzie have proved that the hybrid zero dynamics is existent and
the limit cycle is stable for planar biped robot in paper [14]. If the ankles can
offer enough moment to settle the lateral motion, the lateral motion is
controllable completely and the hybrid zero dynamics of the 3D biped robot is
the same as the planar one. Then the stability of 3D biped robot includes the
stability of periodic motion and the stability of lateral motion.
Stability of periodic motion has been proved in detail in [14]. Here we only
give the conditions to verify the stability of lateral motion. The ground reaction
force applied to the supporting leg is unilateral, so the moment that the standing
ankle can offer is limit. In order to keep the robot lateral stable and not tip
outside, the lateral FRI (Foot Rotation Index) point should in the foot area:

where is the maximal lateral length of the foot.

5. Simulation
Considering the biped robot model with the following parameters (Tablel), we
use MATLAB to simulate the walking process and validate the control strategy.
In order to simplify the simulation, we develop the dynamic model by lumped
mass.
 423

Table 1. Parameters of the robot

torso hip femur tibiae foot

Mass(kg) 10 0 2.5 2.5 1
Length(m) 0.4 0.15 0.3 0.3 0.10
CoM (rn) 0.2 0 0.15 0.15 0.05

The control parameters are a = 0.9 and E = 0.03 . The walking process can
be seen in Fig.3 and the walking parameters as Table2:
Table 2. Walking parameters

2
J,,,, v,, (0-1 V,,,u4x eMm V

0.9872 -64.273 15.103 0.393 1.073

Figure 3. The walking process

Fig.4-7 are some characteristic curve during this wallung process. Fig.4-5
indicate that the motions of the joints are all convergence to stable limit cycles.
Fig.6 indicates that the outputs are complete settled. And Fig.7 indicates that the
lateral motions are stable and the FRI point is in the range of the sole.

6 . Walking Speed Control

Walking velocity V is effected by step length L and step periodT . The gait is
planned by time-invariant strategy, and the walking gesture is control by
geometrical constraints. There is no control to the velocity, so the step period
can’t be controlled directly and the walking velocity would not be the one we
want.
When restricted by N - 1 restrictions, the dynamical model can be reduced
to one dimension zero dynamic. In practice, the reduced model is an inverted
pendulum with alterable length and alterable moment of inertia. The period of
inverted pendulum is determined by the length, moment of inertia and range of
the swing. Based on them, we control the walking velocity by adjust the gesture
of the robot in walking.
424

0.2s , , , , , , , , ,

II

Figure 4. Phase plots of ql

x I 0''

I " " " ' I

0.08
- 0.02

0.0,
I
P 4 0

.O.PI

. -0.02

-1.03

I
0
.
0.1
.
t
.
t.5 1
. .
2 1
.
I
I
IS
I
4
T,rn.,SJ

Figure 6. The output of y6 Figure 7. The lateral FRI trajectory(rn)

To our robot, the length of the pendulum is determined by L, , moment of

inertia is by Sz,,and the range of swing by OM". We can control the walking
velocity by adjusting those parameters.
The fallowing rules can be summarized from experiments:
1. L, and BMm fixed, larger BE,, implies shorter T and larger V .
2. B,:; and& fixed, larger OM" implies longer T and less V .
3. Q
:, and OM" fixed, shorter L, implies shorter T and larger V .
From those rules, we can't get the certain relationship between velocity and
the gesture parameters, but the trend. Using them we can control the velocity by
adjusting it. From those rules, we can't get the certain relationship between
velocity and the gesture parameters, but can get the trend of them. Using them
we can control the velocity by adjusting it.
Here we validate our strategy by simulation. We control the walking
velocity by adjusting 6;:Z0 ,where L, =0.5 and BM" =0.393.P controller is used
here:

e,,N, = el;, +P . - v, ) (1 1)
 425

Where QL,sois the current torso angle, Q[;:o is the last one, VN-' is the last
velocity, and Vd is referenced velocity, P is controller parameter.
Using the controller (1 1) and P = -0.38 , we control the velocity to I d s
first, then accelerate it to 1.3ds, and fix it to 1.15ds at last. Fig.8 is the angle of
torso and Fig.9 is the walking velocity of the process.

Figure 8. Trajectory of e;:& Figure 9. Walking velocity of the robot ( d s )

7. Conclusion
This paper presents a new underactuated 3D Biped Robot. The robot has two
cylindrical sole distributing in lateral plane, which can offer lateral moment to
stabilize the lateral motion. Using the time-invariant algorithm and feedback
controller, the robot realizes 3D dynamical walking. Simulation results indicate
that stable 3D dynamical walking has been achieved. And by adjusting the
gesture of walking, it realizes walking velocity control. Simulation results
indicate that stable 3D dynamical walking has been achieved.

Acknowledgments
This research is supported by National Natural Science Foundation of China (No.
60475035 ).

References
1. Kenji Kaneko, et al, Humanoid Robot HRP-2, Proc, International
Conference on Robotics & Automation, 2004.
2. T. McGeer, Passive dynamic walking, International Journal of Robotics
Research 9 (1990), no. 2,62-82.
3. K. Osuka and K. Kirihara, Motion Analysis and Experiments of Passive
Walking Robot QUARTET 11, Proc. of the 2000 IEEE Int. Conf. on
Robotics & Automation, pp.3052-3056,2000.
426

4. M. Garcia, et al, Efficiency, speed, and scaling of two dimensional passive-

dynamic walking, Dynamics and Stability of Systems 15 (2000), no. 2.
5. Grizzle J W, Abba G, Plestan F. Asymptotically stable walking for biped
robots: analysis via systems with impulse effects, IEEE Transactions on
Automatic Control, 200 1,46 (1) : 5 1 - 64.
6. C. Chevallereau, et al, RABBIT: A Test bed for Advanced Control Theory, PAPER
NUMBER CSM-02-038 Revision 08, June, 2003.
7. Catherine E. Active control of lateral balance in human walking, Journal of
Biomechanics 33 (2000) 1433-1440.
8. J. Maxwell, et al. Mechanical and metabolic requirements for active lateral
stabilization in human walking. Journal of Biomechanics 37 (2004).
9. Steven H. et al. A Three-Dimensional Passive Dynamic Walking Robot
with Two Legs and Knees. The International Journal of Robotics Research
Vol. 20, NO. 7, July 2001, pp. 607-615.
10. Arthur D. Kuo. Stabilization of Lateral Motion in Passive Dynamic
Walking. The International Journal of Robotics Research Vol. 18, No. 9,
September 1999, pp. 917-930.
11. Guobiao Song and Milos Zefian, Underactuated Dynamic Three
Dimensional Bipedal Walking. Proc of the 2006 IEEE International
Conference on Robotics and Automation Orlando, Florida - May 2006.
12. Jessy W. Grizzle, et al. Stable Walking of a 7-DOF Biped Robot. IEEE
Transactions on Robotics and Automation, vol. 19, NO. 4, August 2003.
13. S. P. Bhat and D. S. Bernstein, Continuous finite-time stabilization of the
translational and rotational double integrators, IEEE Trans. Automat. Contr.,
vol. 43, pp. 678-682, May 1998.
14. E.R. Westervelt, J.W. Grizzle, and D. Koditschek. Hybrid zero dynamics of
planar biped walkers, IEEE Transactions on Automatic Control, 2002.
15. C. Chevallereau, A. Formalsky, and D. Djoudi, Tracking of a joint path for
the walking of an underactuated biped, Robotica, vol. 22, pp. 15-28, 2004.
 ROTOPOD: A NOVEL APPROACH
TO EFFICIENT LEGGED LOCOMOTION*

DAMIAN M. LYONS’
Robotics and Computer Vision Lab.
Fordham University. Bronx NY 10458 USA

A number of attempts have been made to integrate the efficiency of wheeled locomotion with the terrain
versatility of legged locomotion, e.g., Univ. Michigan’s Rhex platform and Case Western’s Whegs. Those
platforms cast legs as rotating spokes placed traditionally at the comers of a rectangular platform. In this
paper, we present an alternate approach, with three legs radiating down from a central hub. The energy to
move the platform is generated by a rotating reaction mass mounted at the hub and, at rest, rotating
parallel to the ground plane.
Our approach is to construct a platform whose natural, uncontrolled motion is energy efficient and useful,
and which requires only a small amount of control to produce effective locomotion. We describe the
platform and analyze the characteristics of its uncontrolled motion. Several strategies to produce directed
motion will be presented and evaluated. Simulation models and a prototype platform have been built and
will be discussed.

1. Introduction
Our objective is to develop a legged robot platform that can perform ground surveillance and
reconnaissance activities in an effective and energy-efficient manner. Legged platforms have a
mobility advantage over wheeled and tracked vehicles on rough terrain: They can step into
depressions, onto rises or over obstacles that could defeat a wheeled vehicle. However,
conventional legged platforms offer mobility at a high cost in energy use: Unlike a wheel or
track, a leg must liji as well as propel the robot platform and hence will use more energy.
Wheels, therefore, have an advantage of energy-efficiency over legs. Researchers have
attempted to combine legs and wheels within a single platform, resulting in many different
platform designs. Halme et al.’s WorkPartner [4] has wheels at the end of its legs. Birch et
al.’s cricket-inspired robot [l] has wheels in the front and legs in the back. Saranli et al.’s
RHex [9] has four legs that are single-spokes - legs - mounted on axles, and Quinn et al.’s
Whegs robots [7] have six legs that are each 3 spokes on axles. There is evidence from insect
and associated robotic studies that the reason animals exploit legged locomotion in an energy
efficient manner has as much to do with the use of a tuned physical mechanism as with control
or intelligent planning [8, 101. Our design goal is therefore, to design a mechanism that
naturally produces

This work was supported by STIR P-49411-CI-I1Grant from the US Army Office of research.

427
428

(I) an energy efficient style of locomotion, and

(2) a motion pattern appropriate for ground reconnaissance and sweillance
applications.
In section 2 we introduce a mechanism designed according to this strategy, the rotopod, a
three-legged mechanism that walks by rotating its body around its leg endpoints and its natural
motion is captured as a hypertrochoid. In section 3 we describe how the natural motion can be
controlled. In section 4 we look at the energy efficiency of the mechanism in terms of its
specific resistance. We conclude in section 5.

2. The Rotopod mechanism

Our approach is inspired by the
observation that the RHex and 'virtual' wheel
Whegs leg designs are essentially
one or more wheel spokes (Fig. body
l(a)). An alternate design in which
Leg
legs can be considered analogous to =spoke
wheels is shown in Fig. l(b). In this (a) vertical (b) horizontal
case, the legs are spokes rigidly Figure 1: Legs as Analogs of Wheels
mounted to the frame, as opposed to
mounted on an axle as in Fig. I(@; however, if the platform itself can be made to tilt and
rotate, then the legs will raise and fall to follow the circumference of a 'virtual' wheel as
shown. This mechanism design has the potential to fulfill our design goal because:
(1) it exploits the spokefleg approach already shown to be successful, and
(2) the natural 'rolling' motion of this kind of mechanism can produces convoluted paths
that exhibit a high-degree of area coverage.
Leg motion is effected in Fig. l(a) by rotating the leg axles. However, in Fig. l(b) the legs are
rigidly attached, and the body must be rotated. If the mechanism is extended to include a
rotating reaction mass mounted on the body, then as the mass rotates it can be used to produce
the desired body motions.

Figure 2: Rotopod Mechanism. Figure 3: Mechanism Stepping

 429

Our approach is to investigate whether the natural motion of a tripedal structure surmounted by
a rotating reaction mass (Fig. 2.) can be harnessed to provide a basis for energy-efficient
locomotion for surveillance and reconnaissance activities. We call such a mechanism a
rotopod.

The natural motion of such a platform is to rotate an amount around each of its legs in turn.
The raised legs can be used to step over obstacles that might center-ground a wheeled vehicle
(Fig. 3) giving this platform the mobility to work on a wide range of terrains. When two legs
are raised from the ground, the platform rotates around the third, moving and rotating its
center. The amount of rotation and height of the leg stepping is a function of the leg lengths
and masses and the rotational velocity of the reaction mass.

The resultant epicycloid-like path for the platform has advantages over straight line motion
when the platform must cover an area for search or reconnaissance purposes. The rotating
reaction mass is the principal consumer of energy of in the system; if effective search
locomotion behavior can be obtained with simple control of the natural motion, a highly
energy efficient system will result.

2.1. Descripti5m of the Mechanism

The mechanism concept introduced above raises a number of design parameters. For study
purposes we select the following values for these (see Fig. 4):
(1) Number of Legs: 3. Rational: the minimum number for static stability.
(2) Position of legs on body: evenly distributed around central axis (120' separation)
subtending P=4S0 with vertical midline. Rational: each leg should behave
interchangeably. The value of P is chosen to enhance static stability.
(3) Degrees of freedom per leg: 1 translational degree. Rational This is the minimum
and most direct configuration to change leg length.
(4) Location of reaction mass joint: At apex of legs. Rational: maximize the lever arm
that can be applied by the reaction mass.
(5) Platform lengths and masses: Reaction mass arm limited to about 1 meter max,and
total platform weight limited to less than 20 Kg. Rational: safety and internal testing
430

purposes, though in fact a small reconnaissance drone might be of a roughly similar

size.
An Open Dynamics Engine (ODE) simulation of the 3D kinematics and dynamics of the full
mechanism has been build and used to evaluate motion strategies as well as energy use. This
simulation takes reaction mass motor velocity and saturation torque and leg linear actuator
forces and velocities as input and produces a 3D simulation (e.g., Fig. 3) and various kinematic
and dynamic meas~ementsas output.

A small prototype (-lOcm height) was constructed in [S] that used an alternative rotational,
instead of translational, leg degree of freedom. This ultimately proved limiting. A larger
prototype (-0.45m, Fig. 5) has now been build and is being evaluated. This prototype and the
ODE simulation have the same degrees of freedom, and similar lengths and masses.

2.2. Natural Motion of the Mechanism

If the mechanism is placed on a
level surface, legs fully extended, 1 Poshion of Platform center
and the reaction mass is rotated
slowly ( w cO.2rps for our model)
the mechanism remains stationary. If
the rotational velocity is slowly
increased, then a small chattering
motion of the leg endpoints first
results. Chattering may result in a
x OrtllnaI*
small rotation of the platform around
its center point in the same direction
1 Figure 6: Plot of platform center during natural motion
as the reaction mass rotation. As the
-
velocity is further increased (w 1 rps) the platform tilts in the direction of the reaction mass,
and the chattering turns into the rotation of the platform around each leg endpoint in the
opposite direction to the reaction mass, as the reaction mass passes over that leg. It is this final
mode of motion that we are interested in, and we refer to as the natural motion of the platform,
though the chattering motion may be of some value for small ‘backwards’ motions.

Fig. 6 shows a plot of the (x,y) or plan view of the motion of the center of the platform during
natural motion. Note that it is a series of ‘loops’ which are caused by the tilting of the platform
and the effect of the reaction mass approaching and then leaving the position of the leg
endpoint. This pattern of motion belongs to the family of curves called hypertrochoids
generated by rotating a circle along the inside of the perimeter of a second larger circle. The
standard equations for a hypertrochoid are:

y = (cx-b)sht-hsin x = @ - 6 ) c o $ t + h c . ~ X [ ~ i ’ ) (1)
 431

We can relate the rotopod motion

to these equations. Let 1, be the
length of the leg projection onto
the xy plane, let p be the
Path of center ledmidline angle, and let a be the
max angle that the platform tilts
when the reaction mass passes
over a leg. Fig. 7 shows the plan
view of the mechanism rotating an
E'igure 7: Kinematic simulation of mechanism rotating same
angle around three successive endpoints. amount 0 around each of the three
leg endpoints in turn.As it rotates,
the mechanism center follows a curve with radius r,. The distance moved on each rotation can
be calculated as 4 0 ) as shown, and the relationship with the hyperhochoid curve established
as: B
r, = a - (b+- h) 2& = d(e) = 21eSin/3Sin-2
a - (b- h) = 1,SinP -I,Sin(P a) + (2)

3.Control of Motion
The epicyloid-like path has good ground coverage properties. However, to be useful, the
motion must be controlled to cover a desired area. The natural motion of the platform is fixed

fbl (Cl

(4 (e)
Figure 8: Changin eg Lengths, (a-c) 2 Leg Extensions; ( f ) 1 Leg Extensions.
432

by the leg lengths, masses and rotational velocity, and is, simplifying somewhat, a closed circle
on the ground plain. There are several choices of control parameters include the rotational
velocity, moving the reaction mass along the reaction arm (with a linear actuator) or changing
the leg lengths. We selected changing the leg length because of its energy-efficiency:
( I ) leg lengths can be changed while a leg is in the air, and hence require a minimal
injection of energy into the system.
(2) Once a leg length has changed, the linear actuator model requires no injection of
energy to maintain that length
Fig. 8 shows the effect of modifying leg length, while leaving all other parameters the same:
the radius of the circle of rotation increases in direction proportion. Extending two legs allows
for smaller circles, while extending just one leg generates bigger circles. These graphs were
generated from the ODE simulation. These results mean that the rotopod can be driven along a
curve; however, a practical limitation is that the radius of curvature must always have the same
sign.

The closed circular paths in Fig 8. can be used to build compound motions. There are three
compound motions we have studied, spiral search, epicycloids and area fill. Fig. 9 shows
examples of each of these. Each has the advantage that it has strong area coverage properties,
and is hence a useful mode of locomotion for searching, reconnaissance and surveillance. The
spiral [2] and area till patterns are appropriate when looking at a limited geographical area,
while the epicycloid is a traveling pattern and useful for traversing large areas.

These patterns of motion are very different from the standard approaches to legged or wheeled

(a) Spiral (b) Epicycloid (c) Area Fill

Figure 9: Compound Motion Strategies.
Path of center is shown as dark line. The small circles are leg rotation points. The ground
projections of the three legs are shown in light color.

vehicles [6, 131, and its worth recalling how we got here: The rotopod is designed to have
legged capabilities but high energy efficiency. The patterns shown in Fig. 9 are therefore
energy-efficient search patterns.
 433

4. Energy efficiency
The specific resistance is a common measure of energy efficiency for wheeled and legged
vehicles (though by no means the only one [12]). It is defined as
power
&=
weightxvelocrty
The most efficient legged
locomotion (passive-dynamic
walking [3]) has E between 0.05 ,(,
and 0.08 and the less efficient
legged approaches (e.g., Asimo
biped) can be as great as 1.8.

We used ODE to measure the

mechanical power expended by
the rotopod in natural motion
(e.g., Fig 6). Typical speed for
the robot was about 1.29 d s or
2.8 mph - about walking speed,
and the specific resistance
measured 0.03. Note this
impressively low figure should Figure 10: Potential Efficiency of Rotopod with respect to otner
locornohon
be taken with some caution: It’s
the result of ODE measurements, It’s the only the physical cost of motion, and it includes
motion along the entire penmeter of the hypertrochoid (e.g. 8m in Fig. 6. for a curve of
average radius 0.2m). We have not make measurements on the physical prototypes on the
electrical cost of motion, which is what will determine longevity given current battery
technology. However, this low number, which should be regarded as a best case, does indicate
that the basic idea of the rotopod as energy efficient legged locomotion is a strong one as
shown by the comparisons in Fig. 10.

5. Conclusions
We have described a robot mechanism, the rotopod, that exploits a novel model of legged
locomotion. The natural, uncontrolled motion of the platform has good area coverage
properties and promising energy efficiency. By varying the lengths of the legs of the
mechanism, this natural motion can be controlled and compound motion patterns such as a
spiral (for search), epicycloid (for directed motion) or area fill produced. These results were
obtained using an ODE simulation.

A physical prototype has been constructed that stands approx. 0.45m high and uses one
Firgelli model CYCJ linear actuator per leg and a Korebot/Koremotor embedded computer
running Linux is used as a controller, with wireless 802.11B connection. Preliminary results
434

duplicating those in Fig. 8 have been produced. Compound motions and energy efficiency are
now being evaluated.

Beyond this, the next steps involve adding visual sensing to the prototype by mounting a
camera on one or more legs. The rotating camera can in theory be used to produce a stereo
from motion depth map to aid in foot placement.

References

1. Birch, M., Quinn, R., Hahm, G., Phillips, S., Drennan, B., Fife, A,, Beer, R., Yu, X.,
Garverick, S., Laksanacharoen, S., Pollack, A., Ritzmann, R., A Miniature Hybrid Robot
Propelled by Legs. Proceedings IROS 01, Maui, Hawaii, Oct. 29. - Nov. 3,2001.
2. Burlington, S., Dudek, G., Spiral Search as an Efficient Mobile Robotic Search
Technique. Proceedings of the 16th National Conf. on AI,Orlando Fl.1999.
3. Collins, S. H., Ruina, A. (2005) A bipedal walking robot with efficient and human-
like gait. Proc. IEEE International Conference on Robotics and Automation, Barcelona,
Spain, April 2005
4. Halme, A., Leppanen, I., Montonen, M., Ylonen, S., Robot Motion by Simultaneous
Wheel and Leg Propulsion. 4th Int. Conf. Climbing &Walking Robots (CLAWAR), 2001.
5. Lyons, D., and Pamnany, K., Rotational Legged Locomotion. Proceedings of ICAR
2005, Seattle WA.
6. Mei, Y., Lu, UH., Hu, C., and Lee, CSG. Energy-Efficient Motion Planning for
Mobile Robots. Int. Conf. on Robotics and Automation, New Orleans LA, 2004.
7. Quinn, R. D., Nelson, G.M., Bachmann, R.J., Kingsley, D.A., Offi, J., and Ritzmann,
R. E., (2001). Insect Designs for Improved Robot Mobility. Proc. of Climbing and Walking
Robots Conference (CLAWAROI), Karlsruhe, Germany, pp. 69-76.
8. Quinn, R., Offi, J., Kingsley, D., and Ritzman, R., Improved Mobility Through
Abstracted Biological Principles. Proc. Int. Conf. on Int. Robots and Systems (IROS),
Lausanne Switzerland, Oct. 2002.
9. Saranli, U., Buehler M., and Koditschek, D.E., Design, Modeling and Preliminary
Control of a Compliant Hexapod Robot, IEEE Int. Conf. Robotics and Automation, San
Francisco, California, April 2000.
10. Saranli,U., Buehlerand M., Koditschek D., Hex: A Simple and Highly Mobile
Hexapod Robot. Int. Jour. of Robotics Research, Vol. 20, No. 7, July, 2001, pp. 616 - 631.
11. Schroer, R.T., Boggess, M.J., Bachmann, R.J., Quinn, R.D., and Ritzmann, R.E.
Comparing Cockroach and Whegs Robot Body Motions, IEEE Conference on Robotics
and Automation, New Orleans LA, 2004.
12. Silva, M.F., Machado, JAT, and Lopes, AM. Energy Analysis of Multi-Legged
Locomotion Systems. 4th Int. Conf. on Climbing and Walking Robots (CLAWAR), 2001.
13. Wong, S.C., Middleton, L., MacDonald, B.A., Performance Metrics for Robot
Coverage Tasks. Proc. Australasian Conf. on Robotics and Automation. Auckland, New
Zealand Nov. 2000.
 THE DESIGN OF A HUMANOIDAL BIPED FOR THE
RESEARCH ON THE GAIT PATTERN GENERATORS

P. KRYCZKA
Faculty of Power and Aeronautical Engineering
Warsaw University of Technology
u1. Nowowiejska 24, 00-665 Warsaw, Poland
E-mail: P.Kryczka@onet.eu

C. M. CHEW
Department of Mechanical Engineering
National University of Singapore
10 Kent Ridge Crescent, Singapore 11 9260
E-mail: mpeccm@nus.edu.sg

It was proven that it is possible to build a gait pattern generators, producing

the human-like leg joints trajectories. For this purpose the van der Pol equa-
tions, as well as the neural networks can be applied. A robot described in this
paper will serve as a platform for research on implementation of the mentioned
gait pattern generators in the bipedal antropomorphic structures. The paper
contains the detailed design process description, especially the dynamic anal-
ysis process. The mechanical design methods, as well as the designed preload
mechanism and the foot compliance are also presented.

Keywords: Biped, CPG, CAD/CAE, Compliant foot, Dynamic analysis,

Preload mechanism

1. Introduction
The research on the vertebrae locomotion system, conducted in the 20th
century, led to the conclusion that the rhythmic movements, such as loco-
motion, are generated by the Central Nervous System (CNS).l It was first
suggested2 and then experimentally proven3 that human beings are also
equipped with this kind of system, called the Central Pattern Generator
(CPG). Since 80s the idea of the neural systems generating a set of rhyth-
mic joints trajectories, started being used in robotics. Until now, it was
successfully implemented in various legged robots.
The research being conducted at Warsaw University of Technology aims

435
436

at developing C P G S for ~ ~legged robots. The CPG based on the informa-

tion from the sensors, should be able to produce a set of legs joints trajec-
tories fulfilling the stability rite ria.^ As a first step of the research, a set
of the human gait data was recorded. Based on the recorded data the two
kinds of the gait pattern generators were developed. The first one utilizing
the van der Pol coupled oscillators and the second one the models of neural
networks.s Both approaches proved to be efficient in generating the gait
patterns similar to that of the human being. The current works focus on
building an atropomorphic biped for tests on the implementation of the
created oscillators.
Over the last two decades, a number of advanced humanoidal bipeds
were developed, among the others A ~ i m o HRP-3,1°,~ H7,11 Qrio,12 Ro-
bian,13 Wabian.14 All of these robots have the rigid structure made of metal
or composites. Usually, each joint (DOF) is actuated by the single motor,
unlike in humans, where each joint is actuated by muscle groups. As it is
known from the biomechanics, as well as from our own experience, a human
motion system is very compliant. Our muscles and tendons tend to elongate
under the load, thus making our body more adaptable to the impact loads
(e.g. occurring during the foot initial contact). A progress, worth mention-
ing, towards the more human like structure made the scientists from The
University of Tokyo building a Kotaro15 robot. However, from the robots
mentioned above (except Kotaro) only Robian and H7 are equipped with
the compliant feet.16 Moreover, the compliance in these two robots’ feet
is located in their frontal part, imitating human being’s fingers. This so-
lution improves the ’toe off’ phase of the gait cycle, however it doesn’t
influence the ’initial contact’ phase, which is the source of the disturbing
impact. The solution implemented in the described design influences both
the ’initial contact’ and the ’toe off’ phases of the gait cycle.
The very fast development of the personal computers and their compu-
tational power, in past two decades, drastically improved the design process.
Nowadays, most of the mechanical designers (if not all) use the computer-
aided design (CAD) and engineering (CAE) software as their basic tools at
work. The CAD software lets them for the fast and very accurate modelling
of the structures, whereas CAE software lets them to check the structural
properties (e.g. the stress distribution under complicated load conditions)
of the design and improve its strength and geometrical properties. The de-
sign described in this paper was also performed with use of the CAD/CAE
software, which will be described in more details in the following chapters.
 437

2. The project formulation

The designed robot is meant t o serve as a research platform for CPGs based
motion synthesis mentioned in the introduction. The robot will comprise
only of the lower extremities and the torso. Its mechanical structure will
consist of 12 active DOFs and 2 passive DOFs. The active DOFs will be
directly driven by series of the Robotis' Smart Actuators. The two passive
DOFs will be located in the foot. Justification of their implementation will
be presented in the further part of the paper.
Unlike in the professional design sequence we first had the servos and
then had to suit the biped's structure to their dynamical capabilities. In
order to do that the dynamic analysis had to be performed in the initial
and the final stage of the design process.

3. Dynamic analysis
The main purpose of the analysis was t o set the length of the links, so that
the used servos are able t o actuate the structure. The used DX-113 and DX-
116 Robotis actuators are capable of generating 1 Nm and 2.8 Nm torque,
respectively, By solving a set of inverse dynamic problems for different links
lengths and the gait parameters the joints actuating torque with respect to
the links lengths were found.
The inverse dynamic problem was solved only for the single support
phase, considering robot as a serial kinematic chain and using the Newton-
Euler formulation. l7 The analysis comprised of the following stages:
0 choosing the kinematic model - the kinematic model was limited
to five links (Fig. 2) with the assumption that the pelvis is perpen-
dicular t o the sagittal plane and that only hip pitch, knee pitch and
ankle pitch joints take part in the motion. The five values of the
links lengths were tested 1 = {100,120,130,140,150}[mm],where 1
- length of the thigh and shin links (further referred to as the links
length). The length of the pelvis link was constant in all tests and
was equal 1, = 100 mm.
choosing the gait parameters - the parameters, such as the gait
speed, step length and swing phase time were those of human be-
ing," but appropriately scaled t o suit the size of the robot. The
overall size of the robot was assumed to be 500 mm for all data
sets and accordingly the speed V = 0.076 y , the step length
d = 172 mm and the swing phase duration t = 1.4 s.
0 solving the inverse kinematic problem - the two equations describ-
438

ing the simple geometric relationships (Fig. 1) were used:

+
ryms(a p) +
{ Ic = T l C O S a
y = -rlsina - rzsin(a + p) (1)

For the given pelvis and foot trajectories the set of non-linear equa-
tions (1) was solved numerically with respect to a & p, with use
of the Newton-Raphson method implemented in Matlab environ-
ment. The obtained results were the hip pitch ( a )and knee pitch
(p) angles with respect to time. Finally, from the evaluated data
angular velocities and angular accelerations were calculated.
0 creating a mechanical model and extracting mass properties -
the simplified mechanical model was created in CAD environment.
Because of the biped’s sagittal symmetry only one leg and a pelvis
were modelled. From the created model the mass properties (links’
masses, their centre location and moments of inertia) required for
the dynamic analysis were extracted.
0 solving the inverse dynamic problem - calculation of the torque
in the particular joints was performed in the two, forward and
backward sub-stages, according to the Newton-Euler f0rmu1ation.l~
In the forward sub-stage (from the foot in the support phase t o the
foot in the swing phase), the linear and angular accelerations of the
mass centre were evaluated. Then the forces and moments acting
on the mass centre of each link were calculated, with use of the
following formulae: l7
iflFi+l = mi+i i+ 1 W C , + ~

In the backward sub-stage, beginning from the end of the kinematic

chain (the foot in the swing phase) forces, moments and finally
torque in each join were evaluated, using the following formulae:

‘ n2 .-- iNi + :+lRi+lni+l+ iPcl x iFi + iPi+l x :+lRi+lfi+l

ri = i niT iZi
A

Because of the space limitation, only four graphs of the joint torque for links
length 1 = 100 mm are presented on figure 3. The analysis described above
was preformed with many simplifications. One of them was the trajectory
 439

of the swing leg's foot, in these tests assumed to form a semicircle. The
trajectory of the human's foot of the leg being in the swing phase, during
the normal gait, is much lower, with its beginning and the end almost
tangential t o the support surface. Another simplification was the omission
of the influence of 'toe off' and 'initial contact' phases on the swing phase.
The foot accelerates from the zero velocity and then stops, at the beginning
and the end of the motion respectively. Due to these factors the obtained
joint torque did't fully resemble that of the human being. Nevertheless,
the obtained results were authoritative and sufficient for determining the
maximum torque occurring during the normal gait. Taking into account the
obtained values and the additional structural losses (e.g. friction) the links
length was set to 1 = 100mm.

200

.P
r" 100

50

0
-50 0 50 100 150 2W 250
Dislance[mrn]

Fig. 1. The leg representation. Fig. 2. The stick diagram of the anal-
ysed gait phase.

After completing the robot design, the dynamic properties of the struc-
ture were verified using the A d a m s environment. The A d a m s is a motion
simulation software for analysing the complex behaviour of mechanical
structures. The analysis was performed in the following stages:

0 exporting the 3D model of the robot from CAD environment and

importing it into A d a m s - from the CAD assembly model, cre-
ated during the design process, the individual links were exported
as parasolids (a geometrical model in form of the mathematical
definitions). The exported links were then imported into Adams.
0 assigning the mass properties to the imported model - as with
the parasolid only the geometrical properties of the model were
440

Left Ankle Roll Left K w e Pitch

=fg
z 1.37 . 3 5 ~ 1

1.25 o
-0.5 . : y l
0 50 100 150 200 0 50 100 150 200
Time steps Time steps
Len Hip Pitch Len Hip Roll

9 1.1
P
I-" 0.2

0.1 1.05
0 50 100 150 200 0 50 100 150 200
Time steps Time steps

Fig. 3. Torque in Ankle Roll, Knee Pitch, Hip Pitch, Hip Roll joints obtained from the
analytical dynamic analysis of the leg being in the support phase.

imported] for the dynamic analysis the mass properties had to be

assigned to each link. In this case, it was enough to assign only the
density to the particular solids. The resultant mass, mass centre
location and the moments of inertia of each link were calculated
automatically by Adams.
0 defining the motion parameters of the model - one of the last
steps of this analysis was to assign the motion trajectory to each
of the joints. In order to compare the results of the two methods,
the motion assigned to the joints was the same as the one from the
theoretical analysis, conducted at the initial stage of the design.
obtaining torque in the joints

The results of the dynamic analysis conducted in Adams, for the same
joints as for the initial analysis, are presented on figure 4. This analysis was
conducted with the same simplifications as the initial one. The magnitudes
of torque, except the ankle roll joint, were similar to that conducted before
designing the full structure. During the design the ankle roll axis was shifted
towards the sagittal plane (compared to the model used during the initial
analysis)] which resulted in decreasing the maximum torque in this joint.
The signs of some of the results were opposite compared to that from the
initial analysis, this was caused by the different orientation of the local
coordinate systems in both analysis. The analysis of the structure in the
 441

Adams environment confirmed the initial analytical analysis.

Fig. 4. Torque in Ankle Roll, Knee Fig. 5 . Results of the FEM analysis of
Pitch, Hip Pitch, Hip Roll joints ob- the Ankle Pitch - Ankle Roll link. The
tained from the Adams’ dynamic analy- bright colour denotes the stress concen-
sis of the leg being in the support phase. tration areas.

4. Mechanical Design
The robot structure was designed in ~odell2ngmodule of ~ n ~ g r u ~ h NX4
2cs
- a powerful integrated CADICAMICAE software. As the actuators
used in the design had relatively low torque, making the structure as
light as possible was of the primary importance. In this connection, the
Advanced S z ~ u l u t ~ module
on (utilizing the Finite Elements Method - FEM,
for the structural analysis) was iteratively used to tailor the structure’s mass
and strength properties. An example of results of the FEM analysis of the
ankle pitch - ankle roll link is shown on figure 5 .
The mechanical structure will comprise of the elements cut by laser in
aluminium rectangular pipes. The crucial elements, such as hip-knee and
knee-ankle (Fig. 6) links will form a semi-3D truss, that is known from its
high weight to strength and rigidity ratio.

4.1. Preload mechanism

F’rom the full range of the dynamic analysis results it was visible, that
regardless the length of the links, the load in the ankle roll and the hip
roll joints exceeded the DX-113 actuator’s capabilities. Thus, in order to
442

Fig. 6. The isometric view of the Fig. 7. The front and side view of
knee-ankle joint. the assembly (without the trunk).

decrease the maximum absolute value of the actuator’s load, the pre-load
mechanism was designed.
The pre-load mechanism principle of operation is based on shifting the
range of the asymmetrical load (2), to make it symmetrical (3) and thus
decrease the absolute maximum load.

Let’s take the right hip roll joint as an example. The torque in the
joint, during the motion, ranges from 0 N m to 1.15 N m , where 1.15 N m
exceeds the capability of the DX-113 actuator. By applying the pre-load,
by the torsional spring acting in the opposite direction to the structural
load (Fig. 8), the maximum torque acting on the joint is half of its original
value.

4.2. Foot compliance

The robot’s foot (Fig. 9) comprises of two plates connected in frontal part by
one rotating joint. The rear upper and lower parts of the foot are connected
 443

by a convertible spring. The motion range of the joint is limited, so that

the spring does not fall out in the swing phase of the leg.
The basic principle of operation is that when the leg begins the ’initial
contact’ phase, the rear lower part of the foot touches the ground first. From
that moment the spring is being compressed, until the moment when the
lower part of the foot touches the upper part. Due to this transitional state
between the beginning of the ’initial contact’ and the ’full support’ phase,
there should not be the impact. During the ’toe off’ phase, the compressed
spring begins to release the stored energy and additionally “actuates” the
gait.

Fig. 8. The preload mechanism of the Fig. 9. The isometric view of the com-
right hip roll joint. The upper part of pliant foot.
the structure is not visible.

5. Summary
This paper presents the results of the mechanical model analysis and design
of the robot. It contains a proposed solution to lighten the actuator load by
using a pre-load mechanism. It is expected that the resultant biped robot
will serve as a tool for tests and development of the gait pattern generators
and the influence of the foot passive compliance on the gait stability and
smoothness.

6. Acknowledgements
The authors gratefully acknowledge the support. This work has been sup-
ported by National University of Singapore, and Warsaw University of Tech-
nology (WUT) Research Program. The participation in CLAWAR event
444

was financially supported by Rector of Warsaw University of Technology,

Dean of Faculty of Power a n d Aeronautical Engineering a n d Organizers
of CLAWAR Conference from Nanyang Technological University. T h i s re-
search is supervised by T.Zielinska (WUT).

References
1. C. L. Vaughan, Theories of bipedal walking: an odyssey, Journal of Biome-
chanics 36 (2003).
2. S. Grillner, Neurobiological bases of rhythmic motor acts in vertebrates, Sci-
ence 228,143 (1985).
3. M. R. Dimitrijevic and M. M. Pinter, Gait after spinal cord injury and the
central pattern generator for locomotion, Spinal Cord 37,531 (1999).
4. T. Zielinska, Biological inspirations in robotics: Motion planning, in Proc. of
4th Asian Conf. on Industrial Automation and Robotics, 2005.
5. T . Zielinska, Motion synthesis in walking: Biological and technological as-
pects, in CISM Courses and Lectures, (Springer Verlang, 2004).
6. T. Zielinska, Coupled oscillators utilized as gait rhythm generators of a two-
legged walking machine, Biological Cybernetics 74,263 (1996).
7. M.Vukobratovic, V. Potkonjak and S. Tzafestas, Human and humanoid dy-
namics, Journal of Intelligent and Robotic Systems 41,65 (2004).
8. K. Feng, C. M. Chew, G.-S. Hong and T. Zielinska, Bipedal locomotion con-
trol using a four-compartmental central pattern generator, in IEEE Int. Conf.
on Mechatronics and Automation, 2005.
9. http://asimo.honda.com/.
10. http://www.kawada.co.jp/mechs/hrp-3pIindex.htmlin Japanese.
11. K. Nishiwaki, S. Kagami, J. J. Kuffner, M. Inaba and H. Inoue, Humanoid
JSK-H7: Research platform for autonomous behavior and whole body motion,
in Proc. of fd IARP International Workshop on Humanoid and Human
Friendly Robotics, 2002.
12. http://www.sony.net/sonyinfo/qrio/technology/indexnf.html.
13. F.B. Ouezdou, A. Konno, R. Sellmuti, F. Gravez, Mohamed, B. Bruneau and
0. Bruneau, ROBIAN biped project - a tool for the analysis of the human-
being locomotion system, in Proc. of d h International Conf. on Climbing
and Walking Robots, 2002.
14. http://www.takanishi.mech.waseda.ac.jp/research/wabian/index.htm.
15. http://www.jsk.t.u-tokyo.ac.jp/-ikuo/kotaro/index-e.html.
16. F. B. Ouezdou, S. Alfayad and B. Almasri, Comparison of several kinds of
feet for humanoid robot, in Proc. of sth IEEE-RAS Int. Conf. on Humanoid
Robots, 2005.
17. J. J. Craig, Introduction to Robotics: Machanics and Control (Addison Wesley
Longman, 1989).
18. Z. Bejek, R. Parczai, r. Illys and R. M. Kiss, The influence of walking speed
on gait parameters in healthy people and in patients with osteoarthritis, Knee
Surgery, Sports Raumatology, Arthroscopy 14,612 (2006).
 THINKING ABOUT BOUNDING AND GALLOPING USING
SIMPLE MODELS

KENNETH J. WALDRON’, J. ESTREMERA’, PAUL J. CSONKA’, S.P.N. SINGH’

‘Stanford University, ‘Instituto de Automa‘ticaIndustrial-CSIC, Madrid, ’University of
Western Australia
In this paper simple models are used to develop insights into the mechanics of bounding
and galloping. The focus is on the entire stride cycle, rather than the stance phase. The
models suggest reasons why galloping mammals have their centers of mass closer to the
shoulders than the hips, The galloping model also reveals a hidden symmetry.

Introduction
Here we attempt to understand the important mechanical features of
bounding and galloping using only arguments based on simple models. Despite
their obvious limitations, simple models can lead to deep insights. Based on
those insights, accurate modeling in simulation can then produce reliable
quantitative results.
Front-Rear Force Distribution
It is notable that almost all
mammals, and certainly all cursorial
mammals, are built with their center of
mass closer to the shoulders than to
the hips. Schmiedeler et a1 (2002)
located the center of mass of a large
dog at 65% of the shoulder-hip
distance ahead of the hips. This is
fairly consistent with Jayes and
Alexander (1978) who found that dogs
of various breeds support 61% of their
weight on the front legs. This is a
result of the head and neck being Figure 1: Impulsive equilibrium over the
cantilevered forward of the shoulders duration of a stride for a bound. W is the
in running posture, and also the weight of the system, and D is the drag. 7
is the stride duration. 2IFis the net impulse
forward position of the rib cage and
imparted by the front feet, and 2J, is that
the muscles anchored to it. The reason imparted by the rear feet.
for this persistent feature remains
mysterious. Most likely there are several effects pushing evolution in this
direction.
Experimental evidence (Pandy et al, 1988; Alexander and Jayes, 1983;
Gambaryan, 1974) indicates that when running most animals place their fore
feet well ahead of the vertical projection of the shoulder. The shoulder pole
vaults over the leg thereby converting some horizontal kinetic energy into
vertical kinetic energy and helping to generate the vertical component of the leg
thrust. The net impulse from the front leg winds up having a small rearward

445
446

component in the horizontal direction,

as shown in Figure 1. Thus, the front x
legs brake the system a little, and
provide the largest component of the
Gertical impulse.
In contrast, the rear legs are
placed only a little ahead of the
vertical projections of the hips, or
possibly even further back, and
generate an impulse that is strongly
inclined in the forward direction. The Figure 2: Free body. diagram.
- The center of
horizontal component Overcomes the mass is assumed to be coplanar with the
braking effect of the front legs, and shoulder and hip Joints. The shoulder and hip
joints are assumed to be pairs of intersecting
provides enough forward thrust to revolutes that are, respectively, parallel to the
overcome the drag on the system. The x axis of the body and orthogonal to the plane
vertical component is smaller than that of the leg. The body reference frame is
centered on the center of mass and aligned as
at the front legs, but is sufficient to shown. The principal axes of inertia are
provide moment as assumed to coincide with thex, y, axes.
ensured by the concurrency condition.
Let us first consider a bound. Here the front feet touch down
simultaneously, as do the rear feet. External forces, other than gravity, can only
be applied to the system when the feet are on the ground.
We consider only stably repetitive motion in which the system returns to
exactly the same state after one complete stride, or after one motion cycle. We
assume that angular motions of the body are small. Consequently small angle
approximations can be applied about all three of the attitude axes. This is
certainly not valid for the pitch angle, OY, in a deep bound, as would be used to
leap an obstacle, for example. However, for relatively shallow bounds like those
used for rapid locomotion by some small animals it is a reasonable
approximation. Certainly it is reasonable when galloping, which has the effect
of minimizing pitch rotations.
We also assume that the forces imparted to the body by the legs act
approximately along the leg axes from the point of contact with the ground to
the hipkhoulder center. There is much experimental evidence to back up this
assumption (Pandy et al, 1988; Alexander and Jayes, 1983).
The mass of the legs is neglected and the principal moments of inertia of the
system about its mass center are assumed to be constant. These are the least
defensible of the simplifying assumptions made. As was shown by Schmiedeler
et al. (2002) the leg masses, and variations in the positions of the legs, result in
significant variations in the pitch moment of inertia, Zy.Nevertheless, in the
interests of transparency we choose to regard Zyas constant.
We choose to work in the impulse-momentum domain allowing us to
consider an impulse balance over the entire stride period, T.We also work in a
reference frame fixed to the vehicle body. This does not imply that we regard
the stance period as being instantaneous. Rather, we integrate the contact force
 447

during stance with respect to time. This is analogous to replacing a spatially

distributed force system by its resultant.
If the net impulse imparted to the system by the front feet is 2J,, and that
imparted by the rear feet is 2JR,W and D are the weight of the system and the
drag respectively, the system becomes a three-force system. Therefore the lines
of action of the front and rear impulses must be concurrent with that of the
impulse of the resultant of the weight and drag as shown in Figure 1.
The equivalent dynamic equilibrium equations are:
2JFx + 25, = DT (1)
25, + 25, = !%" (2)
aJFz= bJ, (3)
The effect of the near vertical thrust of the front legs is to pitch the system
back, when viewed in the world reference frame. This allows the rear legs to
contact the ground in a more nearly vertical position than they would have done
without the effect of the front legs. That, in turn, increases the horizontal
component that can be generated by the rear feet without slipping. The rear legs
are free to extend to their limit, in order to generate as much thrust as possible,
without any risk of interference with the fronts,
Impact Energy
It is notable that animals can run
over diverse surfaces without any
noticeable effect on their gait, and at
very nearly the same speed. At first
glance this seems anomalous since the
coefficient of restitution between the
foot and the ground must vary
considerably, and unpredictably. The
reason for this capability is very simple
when one considers the way the foot
works. Figure 3: The KOLT running machine
For simplicity we consider a system that is in operation in the Stanford Robotic
having only one leg, and consisting of a J-~comotionLa~ratoV.
small mass, rn, representing the foot,
connected to a large mass, M , representing the rest of the system by a spring.
Initially the foot is at rest relative to the rest of the system with the spring at its
natural length. The whole assemblage is descending with velocity v,. When the
foot strikes the ground it does not rebound because the mass, M,compresses the
spring, and the spring force holds the foot down. Eventually M is propelled
upwards again by the spring, and energy is added by an actuator assisting the
spring.
The energy that the foot had when it descended is lost, and must be replaced
by the actuator. Because m is very much smaller than M this energy loss can be
accepted. In return, the system can run without adjustment, and with very nearly
the same energetic cost, on most surfaces.
448

The energy that is lost in the impact of the foot with the ground appears to
the system as an increase in the external drag, D. The way this works is
somewhat analogous to the generation of rolling resistance at a wheel. It was
discussed by Schmiedeler and Waldron (1999).
There is a considerable literature on internal energy storage during running,
and the amount of energy stored as strain energy, as opposed to gravitational
potential energy during stance. However, our objective here is to consider the
entire gait cycle, so we will not attempt a more detailed discussion of stance. It
appears that the loss of energy due to the vertical velocity of the foot at impact is
the principal energy drain on the system. Horizontal kinetic energy varies
proportionately relatively little when running fast, and horizontal kinetic energy
is easily converted into gravitational potential energy with minimal loss by the
mechanism of pole vaulting over the leg, discussed above, and the converse is
equally easily and efficiently accomplished by the body falling forward over a
rearward thrusting leg. Impact losses in the horizontal direction can be avoided
by swinging the leg so that the forward velocity of the foot when it touches
down is zero, or at least very small.
Observation of animals running shows that their stride frequency increases
only slowly with speed. In fact, one might hypothesize that the observed
increase in stride frequency is a result of the necessary decrease in stance
duration with speed, and that the effective duration of each ballistic trajectory is
constant. This implies that the vertical velocity at lift-off is for a given foot is
approximately constant, and hence the height of each hop is also constant. Since
the number of footfalls per second is approximately constant, the rate of energy
loss due to the foot impacts is approximately constant. The increase in energy
cost with increasing speed is largely due to increasing external drag. This is
consistent with our observations running the KOLT machine shown on Figure 3
(Estremera and Waldron, 2006, 2007).
The mechanics of the system during stance is often considered to follow the
SLIP model. This has been extensively studied and its properties are well
understood (Raibert and Hodgins, 1991; Saranli et al, 1998). The principal
implication for the present purpose is that the upward vertical velocity
component immediately after stance is exactly equal in magnitude to the
downward vertical velocity
immediately before stance.
We will show that this
condition is excessively
restrictive, and that the slip
model is not, in general,
consistent with a bound, and \xxxx XXXXXXXXNXXXXXXX
cannot be consistent with a (a) (b)
three-dimensional gallop.
Figure 4: Coordinate senses during (a) front foot
Bounding «ta«w anH (M mar frvnt ctanr.p
Let the vertical velocity of the center of mass of the system immediately
before front foot impact be MF; and that immediately after the impact be u'f. The
 449

vertical velocity immediately before rear foot impact is u,, and that immediately
after is u If the vertical impulse delivered by a front foot is J,, and that
IR.

delivered by a rear foot is JzR then applying impulse-momentum across the

impacts gives
ub=uF+- JzF u; = U R +-
JzR
W y ~ = W y 2QJZF
~~-- (4)
M ’ IY
where wyFRis the pitch angular velocity after the front foot stance and before the
rear foot stance, and wyw is the corresponding angular velocity after the rear foot
stance.
Looking at the velocity changes during the two ballistic trajectories (hops)
U R = ub -gzm, U F = Ub - g r , (5)
Here ,z, is the time period from the front foot stance to the rear foot stance, and
, is that between the rear stance and the front. Obviously
,z
zFR+ zRF= T (6)
where T is the stride period.
Applying the condition that the vertical height of the center of mass at the
end of the stride must be identical with that at the beginning
( U; + u R ) z ~+ (U; + uF)zRF = o (7)
M is the total mass of the system and Z, is its pitch moment of inertia about
G. JFzis the vertical component of the impulse delivered by each front foot.
The condition that the system must return to the same pitch angle after a
complete stride gives
F 0
R~ Y R F ~ R =
W ~ F R T F+ (8)
Combining Equations 4 through 8 gives
u; =- (2azm -.(a-b)), U; =- (T(b+a)-2bzm),
2(a+b) 2( a + b)
(9)
U F =- (2azFR -T(U+b)), U R =- ( T ( b- a ) - 2bzm)
2(a+b) 2(a+b)
Similarly, Equations 4 and 8 can be combined to give

where
K=-Mab (11)
I,
Energy Loss
The velocity, vF, with which the front feet strike the ground may be
obtained from Equations 9 with reference to Figure 4:
The energy lost in each stride due to impact with the ground is minimized if
Z F R =T (2b2m,(1-K) + ( U m ~ - b m ~ )+b))/(2(a2mF
(a +b2mR)(I-K))(12)
as the value of zFRthat minimizes the energy consumption. The corresponding
value of U , is given by
450

This might be compared with the corresponding values for a trot for which
TIT is constrained to be 0.5 and there is no body rotation
U , = g2T2(mF+ m,)/16 (14)
The impact energy loss for the bound is always greater than that for the trot.
There is always a term which is a function of K that increases the energy loss for
the bound. K expresses the importance of body pitch rotation.
So far we have ignored motion in the longitudinal direction and analyzed
only dynamic equilibrium in the vertical direction, and rotation about the pitch
axis. However, there is an interchange of horizontal and vertical kinetic energy
resulting from the inclination of the impulses indicated in Figure 1. Usually
there is a rearward component of the impulse from the front feet that causes a
reduction in the horizontal velocity. There is a greater forward component of the
impulse from the rear feet to meet the requirements of Equation 1. Thus there is
an increase in horizontal velocity after rear foot stance.
We assume that the leg is swung so that the longitudinal foot velocity
relative to the ground is close to zero when the foot is placed. This means that
there is no “horizontal impact” causing energy to be lost from the system in the
same way as the vertical impact.
After a complete stride, the total dynamic energy of the system returns to
the same value. Thus a positive value of U,,appears as a corresponding loss of
horizontal kinetic energy at front foot stance, and the same amount of kinetic
energy is returned to horizontal motion at rear foot stance. The work done by the
horizontal component of the impulse is divided between making up the decrease
in kinetic energy due to transfer to
vertical energy, and making up the
energy lost to drag.
Galloping
We follow the suggestion of Minetti
(1998) and regard the gallop as a bound
in which the front and rear leg pairs
“skip”. The advantage of splitting out
the legs at each stance becomes obvious
if we recall that the impulse impelling
each hop is halved. Based on the
discussion above, to a first
approximation this means that the
energy going into vertical motion in
Figure 5: Variations in vertical velocity
each hop is one eighth of that for the and vertical position of center of mass,
hops in a bound, taking into account that and pitch angular velocity and pitch angle
two feet impact at the end of each hop in through a bound stride. The intervals used
a bound. However, since there are four in the text aret, = tR,tw = T - tR.
 451

hops rather than two, the energy loss in a gallop stride should be approximately
one quarter of that in a bound.
We can regard a gallop as being an asymmetric bound in which the front
and rear leg pairs “skip”. That is, we start with a bound and examine the effects
of separating the front feet, and the back feet. Instead of placing the front feet
together, we place foot 1 ahead of foot 2. For a transverse gallop, we would then
place foot 3 ahead of foot 4. The hops FR, and RF of the bound correspond to
the flight phases, 23 and 41 of the gallop, while the skips are the hops 12 and 34.
The stride cycle impulse-momentum relationships are still satisfied. The
separation of the two front foot impulses does not affect Equations 1, 2 and 3
above. Likewise, the relationship between the angular velocities wYz3= wyFRand
wy41= wYw remain the same. The change in angular velocity occurs in two steps
instead of one.
We apply a similar analytical approach to that used for the bound to a
transverse gallop to get, after considerable algebraic development that cannot be
presented here because of space limitations:

u ; = - gz ~ 2 , u z = - u ; , u ; = -
2 2
(-zI2+-
a b 3 4 3= - U k ,

and

wx12 -
2(c + d )
In order to get the above result, it is necessary to assume that the left and
right legs in each pair operate at the same working length, and to apply a
constraint that the roll angle, and angular velocity, w,, at the end of a complete
stride are the same as at the beginning.
On the face of it, this solution, with the flight phases being of equal
duration, is not consistent with the apparent asymmetry of the gait. It is a basic
feature of the gallop that the gathered flight phase (23) is always present, but the
spread flight phase (41) is only present in a fast gallop, and then appears to be
much shorter than the gathered flight phase.
There is a partial explanation if we recall that the stances are of finite
duration and that the impulses are inclined as illustrated in Figure 1. As may be
seen in Figure 8 the front foot touches down far in front of the shoulder, and lifts
off a short distance behind the shoulder. The rear foot touches down almost
directly beneath the hip, and lifts off after the leg has extended far to the rear of
the hip. The effect is to shorten, or erase the period when all feet are physically
out of contact with the ground when going from a rear foot stance to a front foot
stance (41), and to emphasize the flight period when going from a front foot
452

stance to a rear foot stance (23). Thus, the apparent asymmetry between the
gathered flight phase ( 2 3 ) , and the spread flight phase (41) is not necessarily
inconsistent with equal values of z, and ql.Nevertheless, this result does call in
question the assumption that the legs in each pair deliver identical impulses.
Following a development similar to that used for the bound, we can identify
the footfall intervals that minimize impact energy loss:
a~ + b(l - A) u ( i - A) + bK- 1-K-A
TJ34 = T ,223= z~~= T (17)
212 = 2 ( u + b ) ( i - ~ ) 2(u + b)(l - A) 4(1- A)
The corresponding minimal value of the impact energy loss per stride is
U, = g2T ’(mF(
u ( +~A) + b)2 + mR(u + b(ic + A))2)/16(a + b)2 (IS)
It may be observed from Equations 41 that the optimal durations zZ3and z4,
of the flight phases are functions solely of the inertial constants K = Mubll, and
A = Mcdll,. This is important since animals do not increase speed by increasing
stride frequency. Rather, they increase speed by increasing the distance covered
during each stride particularly by increasing the durations of the flight phases.
The stride period, T , declines only slowly with increasing speed. Although we
do not model the stance phase in this paper, we hypothesize that the decrease in
T is due to the decreasing durations of the stance phases. The duration of a
stance phase depends on the angle through which the leg moves relative to the
body while the foot is on the ground. It appears that this does not change much
with changes in horizontal velocity. Hence, the duration of the stance phase is
inversely proportional to horizontal velocity. If the durations of the flight phases
I 2 3 4
remain constant, due to maintenance 1

of constant upward velocities of the

center of mass at lift-off, the overall
stride duration will decline slowly due
to the shortened durations of the
stances.
Equation 18 may be compared to
the corresponding expressions for the
bound and trot given by Equations 13
and 14 above. The energetic
advantage of the gallop over the trot
depends on the value of K + A. The
smaller the value of K + A, the more
attractive it is to gallop from the
energetic perspective, as long as the
inter-stance intervals are close to the
optimal values given by Equations 17.
The significance of Equations 17 and
18 is that the system can increase the
proportion of the stride period devoted Figure 6: Graphical presentation of
variations of velocity and position
to the flight phases, but only at the
variables through the transverse gallop
cost of increasing impact losses. stride.
 453

Conclusions
The models presented here have obvious shortcomings. For example, the
actual duration of each stance period is significant, typically declining from
about 0.2T to 0 . l T with increasing speed (Gambaryan, 1974). This means that
the stances of the front and rear leg pairs actually overlap, creating double
support periods that require higher fidelity models than those presented here.
We have also almost ignored the velocity in the longitudinal direction,
which is, of course, the whole point of running. Likewise, we have not discussed
the effect of splitting the front and rear stances on motion about the yaw axis.
There is a significant effect, but it is easily corrected by a slight displacement of
the center of the front track relative to that of the rear track.
Acknowledgement
The authors acknowledge the support of the National Science Foundation,
Grant Nos. 11s-0208664 and 11s-0535226, and the Secretaria de Estado de
Universidades of the Spanish Ministry of Education and Science.
References
Alexander, R. McN. and Jayes, A. S., 1983, “A dynamic similarity hypothesis
for the gaits of quadrupedal mammals,” J. Zoology (London), 201:135-152.
Estremera, J. and Waldron, K.J., 2006, “Leg Thrust Control for Stabilization of
Dynamic Gaits in a Quadruped Robot, Proc. ROMANSY, Warsaw, Poland.
“

Estremera, J. and Waldron, K.J., 2007, “Thrust Control, Stabilization and

Energetics of a Quadruped Running Robot”, IJRR, under review.
Gambaryan, P.P., 1974, How Mammals Run: Anatomical Adaptions, Wiley.
Jayes, A.S. and Alexander, R.M., 1978, “Mechanics of Locomotion of Dogs and
Sheep”, J. Zoology, London, 185, pp. 289-308.
Minetti, A.E., 1998, “The Biomechanics of Skipping Gaits: a Third Locomotion
Paradigm?”, Proc. Royal Society, B, 265, pp. 1227-1235.
Pandy, M. G., Kumar, V., Waldron, K. J. & Berme, N. 1988, “The Dynamics of
Quadrupedal Locomotion,” J. Biomechanical Engineering, 110, 3, pp. 230-237.
Raibert, M.H. and Hodgins, J. 1991, “Animation of Dynamic Legged
Locomotion,” Computer Graphics, 25, pp. 349-358.
Saranli, U., Schwind, J. Koditschek, D., 1998, “Toward the Control of a Multi-
Jointed, Monopod Runner,” Proc. ICRA, Leuven, Belgium.
Schmiedeler, J.P., Siston, R., and Waldron, K.J., 2002, “The Significance of Leg
Mass in Modeling Quadrupedal Running Gaits,” ROMANSY 14, ed. G.
Bianchi, J.C. Guinot, C. Rzymkowski, Springer, pp481-488.
Schmiedeler, J.P. and Waldron, K.J., “The Effect of Drag on Gait Selection in
Dynamic Quadrupedal Locomotion,” IJRR, 18, 12, 1999, pp. 1224-1234.
Waldron, K.J., 2007, “A Coordination Scheme for an Asymmetrically Running
Quadruped”, to appear Proc. IFToMM World Congress, BesanGon, June.
 USING OPTIMIZATION TECHNIQUES FOR THE DESIGN
AND CONTROL OF FAST BIPEDS

T. LUKSCH* and K. BERNS

Robotics Research Lab, University of Kaiserslautern,
Kaiserslautern, Germany
*E-mail: t.lukschQinformatik.uni-kl.de

K. MOMBAUR* and G. SCHULTZ

I W R , University of Heidelberg
Heidelberg, Germany
*E-mail: katja.mombaurOiwr.uni-heidelberg.de

Fast, dynamic or energy efficient locomotion of bipeds is still an unsolved prob-

lem in robotics. Nature seems to have solved many of the arising difficulties in
thousands of years of evolution, optimizing both mechanics and control along
the way. This paper proposes to use techniques from numerical optimization
and optimal control combined with behavior-based control concepts to address
some of the problems when designing and controlling two-legged robots. Re-
sults from the optimization process will affect both the mechanical construction
and the control strategies.

Keywords: Optimal Control, Behaviour-based Control, Bipeds.

1. Introduction
Speed, efficiency and stability of biological bipedal walking and running is
still unparalleled by contemporary bipedal robots, despite all the technolog-
ical progress made in past decades. The observation of biological motions
leads to the conjecture that nature does apply optimization - not only when
changing the properties of populations in the course of evolution, but also
in the development of characteristic motion patterns of each individual.
One key feature of biological motion is that it fits naturally to the system
itself and to its kinematical and dynamic quantities, and that no un-natural
motion is imposed that would result in inefficiency or increased need for sta-
bility control. Another important characteristic of nature is the exploitation
of elasticity in the passive tissue and in the actuators themselves in order

454
 455

to enhance efficiency and stability of the system.

The goal of the research presented here is to apply the same kind of
optimization that nature applies in the process of designing, building and
controlling a bipedal running robot. A methodology, based on dynamical
multibody system models and efficient numerical optimization techniques
has been developed that assists during all stages of the robot’s construction
and behaviour-based control design - starting at the earliest prototype de-
sign and component selection, and ending only when the robot has acquired
the capability to perform the desired motions efficiently. In this paper, all
steps of the methodology are described and it is shown how it has been ap-
plied to design a leg prototype to be used in a running biped. To control the
robot, a behavior-based reflex network is designed based on results from the
optimization and knowledge on neuro-motor control in nature. Optimiza-
tion is also used to increase the exploitation of elasticity in the system by
adequately designing springs for the running motion. In current research,
this same methodology is to be applied in the design and construction of
the full biped robot.
Other projects focusing on dynamic locomotion of bipeds include the re-
search on the robot Lucy, where elasticity is included in form of pneumatic
muscles [1,2]. Another project working on dynamic walking and running
is the Rabbit robot. Questions examined include stability of posture or
robustness against external forces [3]. The benefits of self-stabilizing sys-
tems are exploited in the field of passive-dynamic walking. With robots like
Denise or Meta from the University of Delft it is examined how actuation
can be included to this concept [4,5].

2. Methodology: Optimization-based design and control of

robots and components
Inspired by nature, a methodology for designing and controlling robots
and robot components based on dynamic models of the robotics system
and efficient state-of-the-art optimization techniques has been developed.
The proposed methodology requires a frequent interaction and exchange of
results between the computational and the experimental side. The following
steps are taken:

(1) Specify a preliminary mechanical design, including dynamical proper-

ties of material, include ranges of characteristics of motors and other
components. Define the desired form of locomotion (i.e. walking, run-
ning, hopping, type of ground etc.)
456

(2) Create a dynamic model for this initial construction and the desired
locomotion, include all realistic constraints in the model
(3) Optimize this model with design and control parameters as free vari-
ables
(4) Modify design according to results from optimization
(5) Iterate 2-4 if necessary
(6) Integrate optimal control results into a generalizing behavior-based con-
trol scheme
(7) Build the prototype and implement the control
(8) Validate by comparing the results from the prototype with those from
the simulated model

3. Experimental Setup
As described in the previous section, the first step of the proposed method-
ology is to develop an initial design for the robot. To verify the feasibility of
the methodology in the framework of this initial study, instead of directly
investigating a full biped, a single, jumping leg is started with. Jumping on
the spot is an extremely dynamic, periodic movement putting high demands
on the mechatronical design and the control, this way providing meaningful
evidence on the quality of the proposed method.
The experimental setup of the prototype leg on the treadmill is shown
in figure l(b). The leg has an active hip and knee joint with one rotary
degree of freedom each. Being fixed to a linear vertical slider, it can perform
vertical motions in the hip. Each of the joints is powered by a non-retardant
drive composed of a d/c motor with low ratio gear allowing for a freely
moving joint and thus for less mechanical stress and more time to react to
disturbances.
The knee joint is supported by a spring mounted in parallel to the drive
with fixed equilibrium position. This newly designed pneumatic, rotatory
spring (see fig. l ( a ) ) has the advantage that it can produce high torques and
the stiffness can be adjusted by changing the working volume using switch-
able separations including valves (for more details, see [S]). A custom-made
load cell in the lower limb allows measuring of the force in leg length direc-
tion (z-axis) and the torques around the x- and y-axis during impact and
stance phase. The overall weight of the prototype sums up at about 12kg.
This initial construction forms the starting point for the subsequent opti-
mization process. Several of the mechanical design parameters are defined
as free variables for the optimization.
 457

Fig. 1. (a) Design of the pneiiriiatic spring showing the moving piston and one separa-
tion - (b) The initial single leg prototype mounted on a linear slider of the test rig.

4. Mathematical Model, Optimal Control Problem

Formulation and Solution
Set,ting up a mathemasticalmodel of the leg is step (2) of the above inethod-
ology. A rigid multibody system model of the prototype leg has been estab-
lished including all free design parameters p and free input quantities u(t)
driving the system (in this case the torques produced at hip and knee). The
model of this leg performing a running - or rather hopping - motion on this
leg results in a hybrid dynarnical system involving both continuous phases
(flight and contact) and discrete model phases (discontinuous touchdown
event).
Clioosing the three coordinates hip height y , absolute angle of upper
leg au, and relative angle of lower leg $it we obtain a set of differential
equations of form

q t ) = f i @ ,4 t ) ,? ~ ( t ) l p )for t E [ m f t o f f , ~tozlchdown] (1)

for the flight phase with zT = (y?a%, (61, .jr,&, &)" and differential alge-
braic equations for the contact phase

k(t>= ~ ~ ( ~ , ~ ( ~ ) ? ~ ( t ) , p ) (2)
0 = d t , 4 t ) l 4 t ) ,P ) for t E b-touchdouln37 2 i f t o f f l (3)
The touchdown discontinuity is described by
-1-
Z(%7tlchdolun) = J(4'1ilS,chdown)1 P>* (4)
458

Starting points rj of new phases do generally not explicitly depend on time,

but are implicitly defined by the roots of switching functions

sj(t,
)j.(. ,P) = 0. (5)
The periodic hopping cycle considered in the model starts in the flight
phase at the highest point, goes through the rest of of the flight phase, the
touchdown discontinuity, contact phase, and then the first part of the flight
phase until the cycle is completed. Periodicity constraints are formulated
) z(T),
on all state variables ~ ( 0 = where the cycle time T and all phase
times are free variables. Realistic constraints are imposed on all design
parameters, controls and states. Step (3) of the methodology consist in
determining optimal motions for this model. This results in solving the
following multi-phase optimal control problem:
rT

s. t. model eqns. (1)/(2) and (4)

+ addl. equality and inequality constraints
Problems of this form can efficiently be solved by optimization techniques
based on the direct boundary value problem approach using multiple shoot-
ing [7-91. This technique involves a discretization of both state and control
variables transforming the optimal control problem into a nonlinear pro-
gramming problem (NLP) with simultaneous simulation for the evaluation
of objective function and constraints. For an application of these methods
to hybrid gait problems and required extensions if stability is used in the
objective function, see [lo].

5. Optimization Results for Prototype Leg

The goals of the optimization runs performed for the prototype leg were
0 to determine if the original leg design and the initial choice of compo-
nents would be adequate to preform the desired hopping motion (with
a desired hopping height of 20-30 cm, and at least 10 cm)
if this was not the case, to establish a better setup, parameters and
actuation patterns
0 to identify which modifications are most crucial in improving the per-
formance of the leg
0 to evaluate the use and the best design of the pneumatic rotary spring
in order to exploit spring elasticity to produce more efficient hopping.
 459

Since the prototype leg only moves on the the test rig, stability has not
been considered in the optimization at this point. The optimal control prob-
lem formulations aimed at finding periodic hopping motions that maximize
jumping height under different conditions. In a first series of optimization
runs we have determined the dependency of the maximum hopping height
on the average torque limits and the peak torque limits in the knee and the
hip for a first set of design parameters of the prototype leg. The results in
figure 2 show that an average torque limit of 20-30 Nm and a peak torque
limit of 40 - 50 Nm are required. For the choice of a maximum average

06
06

$ ":
I f GI
-5 vr $m
01 03
at

.n"s.(o4Y*MB"I
"5
p a * -t YnH ,"*
Fig. 2. Dependency of maximum hopping height on limits of (a) average torques and
(b) peak torques

Fig. 3. Dependency of maximum hopping height on the maximum (a) stretching and
(b) bending angle at the knee

torque of 20 Nm we investigated the influence of several parameters on the

maximum hopping height:
0 the maximum stretching angle of the leg: the more the leg can be
stretched, the higher is the jump, but the possibility to overstretch
the leg at the knee only leads to slight increases of the hopping height,
see figure 3
0 the maximum bending angle of the leg: this has a significant influence,
see figure 3, should be at least 100 deg, more than 160 deg is not useful
460

0 the maximum pressure in the pneumatic spring has a major influence on

the hopping height (see figure 4). In this case, three different average
torques have been investigated, and we found that with a maximum
pressure of 30 bar, even an average torque of 13.6 Nm was sufficient to
produce a jumping height of 25 cm. Pressures of .... in the spring are
realistic.

"_
rnlXlrnYrn preS.ws I" pmumauc spring pa,l

Fig. 4. Dependency of maximum hopping height on the maximum pressure in the pneu-
matic spring

In addition, we have investigated the effect of a treadmill on the hopping

height. Not too surprisingly, having to produce an additional forward or
backward motion while joint torques are subject to the same constraints,
results in a reduced hopping height. With an maximum average torque of
20 Nm, the prototype leg only is able to lift off for treadmill speeds up to
M 1.2mfs

After all these initial computations that served to evaluate the original
prototype design some necessary modifications have become obvious (com-
pare step (4) of the methodology) such as the introduction of gears with
higher ratios in hip and knee in order to be able to produce the high torques
required for the desired motions. These resulted in additional masses that
were taken into account in a new set of computations performed subse-
quently. With limits of average torques at 27.2Nm and 17.ONm for knee
and hip, and limits of peak torques at a factor 2.5 respectively a maximiza-
tion of jumping height resulted in the state variable and control histories
shown in figure 5. The jumping height was 13.5cm. The resulting parame-
ters of the optimally designed pneumatic spring, in which it best supports
the computed target motion, were 1.88bar initial pressure, a compartment
partition at 185 deg, and a spring zero position at -12.5 deg.
 46 1

nip anglr MII

*
0

0 C.2 0.4 0.6 0.8

Fig. 5 . (a) Position and velocity variables and torques of optimal motion

6. Behavior-based Control of Prototype and Experimental

Results
After optimizing the initial construction in several iterations, a suitable
control must be conceived (step ( 6 ) of the methodology). Beside the opti-
mal mechanical parameters, the optimization process also provides optimal
control trajectories for the motion aimed at in the undisturbed case. Those
trajectories can directly be used as open-loop control strategy in hope the
real robot does not differ to much from the modeled one. This way a one-
on-one comparison between the optimal control results and real prototype
is possible.
The second strategy is to derive a reflex control. The optimal control
trajectories are analyzed to find generalized reflexes that produce similar
control output in the undisturbed case, but also give reasonable results
with arising external disturbances. The selection of reflexes is also influence
by knowledge on neural control from biology. This control is realized as
462

behaviour-based network.
In previous works, a framework for the development of behaviour-based
robot control has been designed that has been successfully used on many
different machines (see e.g. [ll]).It consists of a network of reflexes or be-
haviors. Besides the sensor input Zthat is transformed to the control values
13by the transfer function F , each behavior has a special motivation input
L and two meta information outputs a and r that state the current activity

and target rating and are used for behavior-coordination with fusion nodes.
The basic behavior module is shown in figure 6.

a r I

e Y

Fig. 6. The basic behavior module with its interface

Figure 7 shows the resulting reflex network. The set posture reflex guar-
antees as slightly flexed leg stance for stable standing and secure landing
after the flight phase. During ground contact the hip torque is reduced
by the relax hip reflex similar to what can be seen in the optimal control
trajectories. This also reduces mechanical restraint. To achieve enough ac-
celeration to take off, the knee joint has to be bend sufficiently, which is
done by the prepare j u m p reflex after touching the ground. Finally, the push
off reflex produces full torque in the knee and hip joint as soon as the knee
is bended far enough to produce maximum acceleration. The control flow,
the fusion nodes and the inhibiting connections can be taken from figure 7.
As the finishing of the real prototype leg is still pending, the behavior-
based control is tested in a dynamics simulation environment. The activities
of the reflexes during one jumping cycle can be seen in figure 8. Despite no
explicit cyclic behavior like an oscillator being present, the leg shows peri-
odic jumping with nearly exactly the same cycle duration as was calculated
as optimal by the optimization process. The control shows robust behavior
to disturbances, e.g: sudden changes to the ground height.
Figure 9 shows the resulting joint angles and the upper body posi-
tion. The ground contact force is also plotted to show the points in time
when the leg is hitting the ground, is slightly bouncing back, resting on the
 463

-
I
I
Hardware Abstadion Layer
t

(Prototype) I Simulation
I
1 I- control flow

Fig. 7. The reflex network producing robust jumping of the prototype leg. Edges ending
in a filled dot represent inhibiting connections. The rounded boxes denoted with an f
are fusion nodes (e.g. f: knee torque).

1.
0.8 -
$
B 04
~~~~
06 0.6 .
50 04 04 0.4 .
r
a 02
~~~ 02 02 0.2
0-
0
0 0 2 0 4 06 0 6 0 02 04 06 08 0 02 0 4 06 0 8
1 w4 1I s e l 1-1

Fig. 8. Activity of the reflexes during one jump cycle

ground and finally lifting off. The resulting trajectories are similar to those
predicted by the optimization calculations. The pneumatic spring always
produces a certain torque as the equilibrium position is set to an negative
angle and the spring is pre-stretched. The slider position shows a lift-off
height of about 5cm.

7. Extensions of Methodology to Bipedal Running

Our ultimate goal is to apply the described methodology to construct a
bipedal running biped. The use of the prototype leg in this context was
twofold:
0 t o test the methodology of optimization-based design
0 to manufacture a first component that might serve as a basis for the
future bipedal robot.
 0 -
p 02-
04.
08
f 0 8
1-

f lsecl f w f IW t

Joint angles, position of upper body and ground contact force during one jump

The target configuration of the running biped has nine bodies (trunk,
shanks, thighs, feet and arms) resulting in 11 degrees of freedom if mo-
tions are restricted to the sagittal plane. We have already established the
bipedal model and produced first optimization results for energy-efficient
running at different speeds, using human-like geometries and mass distri-
butions. One of these optimized motions is shown in figure 10.
In our computations, all model parameter ranges as well as torque limits
etc. will be adapted to the technical restrictions of robotics components as
soon as we enter the design phase of this robot.

Fig. 10. Optimization of running biped

For the computations and the design of the bipedal running robot, sta-
bility and robustness of the motion play a crucial roles, in addition to
the goal of moving efficiently at certain speeds. We therefore also include
open-loop stability and robustness as optimization criteria for the biped
comput ations.
Our latest biped optimizations have shown that it is possible to achieve
stable running without feedback for human-like model parameters, and we
therefor expect that we will also be able to heavily exploit natural stability
by correctly designing the biped. Detailed descriptions of the bipedal robot
model and the optimization results are given in two other publications
currently in preparation.
 465

8. Outlook
After validating the methodology with the prototype leg, the concept will be
continually extended to two-legged systems. The design of a reflex network
for a biped robot has already been started. The development and control
of biped prototypes with increasing complexity including elastic elements
will be the next steps.

9. Acknowledgements
The research project has been supported by the German Ministry of Educa-
tion and Research (BMBF) within the framework "Bionik - Innovationen
aus der Natur".

References
1. B. Verrelst, R. Van Ham, B. Vanderborght, F. Daerden and D. Lefeber,
Autonomous Robots 18, pp. 201 (2005).
2. J . Vermeulen, D. Lefeber, B. Verrelst and B. Vanderborght, Trajectory plan-
ning for the walking biped lucy, in 7th International Conference on Climbing
and Walking Robots, 2004.
3. C. Sabourin, 0. Bruneau and G. Buche, Experimental validation of a robust
control strategy for the robot rabbit, in IEEE Conference on Robotics and
Automation (ICRA), 2005.
4. M. Wisse, D. G. E. Hobbelen, R. J. J. Rotteveel, S. 0. Anderson and G. J.
Zeglin, Ankle springs instead of arc-shaped feet for passive dynamic walkers,
in IEEE International conference on Humanoid Robots, 2006.
5. S. Collins, A. Ruina, R. Tedrake and M. Wisse, Science 307,pp. 1082 (2005).
6. T. Luksch, K . Berns and F. Florchinger, Actuation system and control con-
cept for a running biped, in Fast Motions in Biomechanics and Robotics, eds.
M. Diehl and K. Mombaue (Springer Verlag, 2005).
7. H. G . Bock and K.-J. Plitt, A multiple shooting algorithm for direct solution
of optimal control problems, in Proceedings of the 9th IFAC World Congress,
Budapest, 1984.
8. D. B. Leineweber, I. Bauer, H. G. Bock and J. P. Schloder, Comput. Chem.
Engng 27, 157 (2003).
9. D. B. Leineweber, A. Schafer, H. G. Bock and J. P. Schloder, Comput. Chem.
Engng 27, 167 (2003).
10. K . Mombaur, H. Bock, J. Schloder and R. Longman, Open-loop stable solu-
tion of periodic optimal control problems, in Z A M M , 2005.
11. J. Albiez, T. Luksch, K. Berns and R. Dillmann, The International Journal
on Robotics Research, Sage Publications vol. 22,pp. 203 (2003).
Using Virtual Model Control and Genetic Algorithm to Obtain
Stable Bipedal Walking Gait Through Optimizing the Ankle
Torque

Van-Huan Dau, Chee-Meng Chew and Aun-Neow Po0

Department of Mechanical Engineering, National University of Singapore,
10 Kent Ridge Crescent, 1 1 9260, Singapore
E-mail: dvhuan, mpeccm, mpepooan@nus.edu.sg
www.nus.edu.sg

This paper presents a new approach of controlling the biped robot to achieve
stable dynamic walking by optimizing the stance-leg ankle behavior. In this
study, the motion control in the sagittal plane of the robot is formulated using
a control framework known as Virtual Model Control. Genetic Algorithm is
used to search for optimal gains of the controllers such that stable walking
gait with smooth velocity profile can be achieved. The stability of the biped
is checked using the zero-moment-point criterion. Simulation results show that
the obtained walking gait is stable and its forward velocity profile is very
smooth.

Keywords: Bipedal walking, Virtual Model Control, Genetic Algorithms

1. Introduction
Stability is probably the most important issue in bipedal walking. Although
many approaches have been introduced to tackle this problem, this is still
an open question. One of the key characteristics of biped robot that make
it difficult to be controlled is that bipeds are non-linear, high dimensional
systems. Another important factor that causes most of the difficulties in
maintaining stability is the underactuation between the feet and the ground.
Human beings can walk stably in different kind of terrains because they can
flexibly control the ankle joints and thus effectively control the contact be-
tween the feet and the ground. For robots, however, controlling the ankle
joint to obtain stable walking is much more challenging. M. Guihard et al.
(2002) introduced the notion of hip and ankle strategies coming from phys-
iological observation on the human, in the control and design of a biped
robot. K. Young Yi (1997) designed a biped which has a compliant ankle

466
 467

joint. The compliant joint gives a flexibility of joint and compliance when
touching the ground thus making good contact between the feet and the
ground. However, no research so far focuses on optimizing the ankle behav-
ior to achieve a smooth, human-like walking and to regulate the walking
speed of the biped.
In this paper, we propose a method in which Virtual Model Control
(VMC)1$2is used as the control framework and ankle torque of the stance
leg is formulated and optimized such that stable walking gait could be
achieved with a smooth velocity profile and the ankle torque is as low as
possible. Genetic Algorithm (GA)4 is used as a searching engine t o search
for the optimal value of the key parameters which optimize the ankle torque.
Zero-moment-point (ZMP)5 has been widely used as a stability criterion in
bipedal walking, hence it will be used to verify the feasibility of the biped’s
walking gait in this study.
This paper is organized as follows. Section two briefly describes NUS-
BIP, the biped being simulated. Section three shows the implementation of
Virtual Model Control and explain the importance of controlling the stance
ankle t o achieve stable walking. In section four, the application of GA to
search for the optimal value of ankle gains will be presented. Section five
is devoted t o walking algorithm and ZMP computation. Simulation results
and discussion will be shown in section six. Finally, some conclusions will
be made in section seven.

2. The Biped Robot

The robot considered in this study is a 7-link biped robot named NUSBIP,
which was built and developed in our lab (see Fig. 1).The total weight is
about 22.2 kg. The biped has 12 DOFs. Each leg has six active DOFs of
which three DOFs at the hip (pitch, yaw, roll), one at the knee (pitch) and
two at the ankle joint (pitch, roll). Each DOF is driven by a DC motor and
integrated with an angular position sensor t o measure the relative angle
between two consecutive links. Each of the feet is equipped with four force
sensors (two at the toe and two at the heel) t o sense the contact forces
between the feet and the ground.

3. V i r t u a l Model Control ( V M C ) I m p l e m e n t a t i o n
Virtual model control is a motion control framework using virtual com-
ponents to create desired joint torques. These virtual components such
as springs, dampers, dashpots, masses, or any imaginable components are
468

Fig. 1. Picture of NUSBIP

placed at strategic positions to achieve certain task. Virtual forces gener-

ated by virtual components will act on the robot as if they really exist.
Then these forces will be transformed into joint torques using the relation
7- = J T F to control the robot as desired. For more details on VMC, please
refer to Ref. 1. Fig. 2 shows the mounting of virtual components on the
body, the stance and the swing leg.
For the stance leg, the torques of the ankle, knee and hip joints
[7i, 7-k rhIT can be derived from the virtual forces [ f z fi which are
generated by the virtual components, by means of Jacobian:

Where f z , fi and fe are respectively the horizontal, vertical and rotational

virtual forces applied on the body of the biped, Ca = COS(Cpa); ca+k =
+ +
cos(cpa vk); sa = sin(cpa); s,+k = sin(cpa cprc).
The effect of the virtual force f z is to pull the biped to move forward
while f i and f e to maintain the body’s height and pitch angle, respectively.
Ideally, if the stance foot is fixed to the ground all the virtual forces can be
set independently to achieve a desired gait. However, in humanoid robot the
contact between the foot and the ground is unilateral and under-actuated.
As a result, if the resultant ankle torque derived from the prescribed virtual
forces is too high, the foot may be rotated, and this may lead to instability.
In addition, when the motion of the swing foot is specified to meet the step
length and speed requirement and the trunk is kept upright as in human
 469

Fig. 2. The placement of virtual components on the biped.

walking, t o obtain stable walking the ankle torque must be constrained. As

the ankle torque is constrained, the virtual forces cannot be set arbitrarily.
The ankle torque is formulated as follows:
7, = Kp(Vdesired - V ) + K d ( 6 d e s i r e d - 6) (2)
Where v is current forward velocity of the body, Vdesired is the desired
velocity. The optimal value of gains K p and K d will be searched by GA.
Since ankle torque is constrained and fz, f e are set independently to satisfy
the desired hip’s height and hip’s pitch respectively, fz must be solved from
(1). Once f z is known, the knee torque Tk and hip torque ~h are determined
as in equations (3) and (4):
C
rk = -7,
A + (D- -)BC
A fz f
C
(-A - 1)fe (3)

r h = -fO (4)
For the knee torque in (3) t o be determined, the denominator A = -Llc, -
L2C,+k must be non-zero. A equals to zero when Pk +
2q7, = 180’ (for the
case of L1 = L 2 ) . This condition only happens when the hip is at the same
level as the ankle joint. However, this condition may not happen during
walking, so we are sure that A is always different from zero.
For the swing leg, the torque of the joints is easier to control because all
the virtual forces can be set independently. The Jacobian for the swing leg
is derived from the swing foot t o the hip. The joint torques are determined
using the same relation r’ = ( J ’ ) T F .
470

The parameters of the virtual components are spring stiffness (kz,kz),

damping coefficient ( b z , b z ) , torsional spring stiffness (ka) and torsional
damping coefficient (ba).

4. Genetic Algorithm Implementation

4.1. GA’s parameters
In this study, GA is used to search for the optimal value of ankle gains K p
and Kd in ( 2 ) . Each gain is coded into a string of ten-bit binary number.
However, we need to search for both gains simultaneously. In order to do
that, we combined the strings of K p and Kd into one single string of twenty-
bit binary number. All the genetic operations (reproduction, crossover and
mutation) will be performed on this string. The probability of crossover is
selected to be 80%. The mutation probability is 2%.

4.2. Fitness function

The objective that we want to achieve is reflected through the fitness value.
In this study, the interested parameters are the ankle gains of the ankle
torques. These parameters will be searched by GA to find the optimal
values satisfying the requirement that the horizontal velocity profile of the
body is smooth, the average velocity is close to desired velocity and the
ankle torque is minimized. The performance index is as followed:
. n

Where

Where n is the number of walking steps, m is the number of integration

steps and i denotes ith step.,,V and Vmin are respectively the maximum
and minimum speed within that step. Vavgis average speed of the biped
which is equal to distance of walking divided by duration of walking.
is desired speed of the biped. w1,w2 and w3 are the weighting factors. Tref
is the reference torque value.
The performance index must be minimized. However, as a convention,
in GA the fitness function is always being maximized. Therefore, the fitness
function is F = +.
 471

5 . Walking Gait
5.1. Walking Gait Generation
In this approach, thanks t o the use of virtual model control, the walking
gait is generated using simple, intuitive strategies instead of planning foot
and hip trajectories. The advantage of this method is that it is simple,
intuitive. The strategy adopted for the swing leg is as follows:
(i) Lift the swing foot t o a specified height H,. This job can be achieved
easily by setting the equilibrium point for the vertical spring-damper.
(ii) Swing the leg forward with a desired speed t o a specific horizontal
end-position (this end-position is determined such that desired step
length could be achieved). In this case, the spring setpoint is moving
with the same desired speed.
(iii) Land down the swing foot while maintaining the foot angle such that
the foot is parallel to the ground. This can be done by damping the
swing foot in 5 direction.

5.2. Stability Verification

In this method, the ZMP5 criterion is used to check for stability of the biped.
It is noted that no desired ZMP trajectory is proposed in this approach and
ZMP is only used to verify that the achieved walking gait is feasible. For
the computation of ZMP position please refer to Ref. 3. The biped is stable
if the computed ZMP is within the support polygon.

6. Simulation Results And Discussions

The specifications of the simulated biped are taken from a real biped, which
was named NUSBIP and developed in our lab (Fig. 1).Table 1 summarizes
its specifications and desired walking parameters. The simulation is done
in Yobotics, a dynamic simulation software which allows the running of
batches of simulation. Velocity profile of the biped when the ankle gains

Table 1. Specifications of the simulated biped.

I Parameters I Value I Parameters I Value I Parameters I Value I
Body mass 10.45 kg Thigh length LZ 32 cm Body height 65 cm
Thigh mass 2.68 kg Shank length L1 32 cm Step length 35 cm
Shank mass 1.68 kg Foot length 23 cm Desired speed 0.75 m/s
Foot mass 1.52 kg Foot thickness 5.5 cm Desired acceleration 0 m/s2

are optimal values ( K p= 13.13, Kd = 1.06) is shown in Fig. 3. From the

472

figure we can see that the velocity profile is smooth and the average velocity
is very close to 0.75 m/s which is the desired velocity of the biped. Fig. 4
shows the torque applied at the right ankle. It is observed that the maximum
ankle torque is about 20 Nm. This maximum torque only appears at some
time instants, most of the time the ankle toque is below 10 Nm which is
quite small.
To assure the effectiveness of the proposed method we have to prove
that the achieved walking gait is stable. In view of this, we use ZMP as a
criterion to check for stability. The biped is stable if the computed ZMP falls
within the foot support polygon. Since we consider only the sagittal plane,
the support polygon becomes the upper and lower bounds for the ZMP. In
single support phase, the lower bound is the heel and the upper bound is
the toe of the supporting foot, respectively. In double support phase, the
lower bound is the heel of the back foot and the upper bound is the toe
of the front foot. It is shown in Fig. 5 that the ZMP trajectory is always
within the lower and upper bound. Therefore, the resulting walking motion
is stable. This result convinces that the stance-ankle is very important to
achieve a stable walking motion especially when the swing leg motion has
been chosen to obtain certain desired tasks. Fig. 6 shows the stick diagram
of the biped walking on level ground, the solid line represents the right leg
and dashed line represents the left leg.

6 6.5 7 7.5 8 8.5 9 9.5 10

Time (sec)

Fig. 3. Velocity profile of the biped when the ankle gains are optimal.

i f 0 5 10 15 20 25 30
Time (s)

Fig. 4. The torques applied at the right ankle

 473

2- r&
A ._ -. .I

0.5 1 1.5 2 2.5 3

Time (sec)

Fig. 5 . ZMP trajectory of the biped

0.8

-- 0.6

0.4
-m
= 0.2
Y -
0 0.5 1 1.5 2 2.5 3
Walking distance (m)

Fig. 6. The stick diagram of the biped depicted at 0.1s apart

7. Conclusion
In this paper, we have presented a method in which VMC is used as a
control framework and GA is used to search for optimal value of ankle
gains. The simulation results show that the optimal values of the ankle
gains result in a smooth and stable motion. The average speed is also close
to desired speed. Our study shows that when the motion of the swing
leg is pre-defined, appropriately controlling the stance ankle torque is an
effective way t o achieve a stable and smooth walking motion. In this study,
the generation of the walking gait is very simple and intuitive thanks to the
use of Virtual Model Control.

References
1. J. Pratt, CM Chew, A. Torres, P. DilWorth, F. Pratt, Virtual Model Control:
An Intuitive Approach For Bipedal Locomotion, The International Journal
Of Robotics Research, Vol. 20 No. 2 February 2001, pp 129-143.
474

2. CM Chew, G. A. Pratt, Dynamic bipedal walking assisted by learning, Robot-

ica , (2000), Vol. 2 0 , pp. 477-491.
3. C. Qiang Huang, Kazuhito Yokoi et al., Planning Walking Patterns for a
Biped Robot, IEEE Transactions on Robotics and Automation, Vol. 17,No.
3, June 2001.
4. D.E. Goldberg, Genetic Algorithm in Search Optimization and Machine
Learning (Addison-Wesley, MA), 1989.
5 . Vukobratovic, B. Borovac, D. Surla, and D. Stokic. Scinetific Fundamentals
of Robotics '7. Biped Locomotion: Dynamics Stability, Control and Application
(Springer- Verlag, New York), 1990
6. Keon Young Yi, Walking of a Biped Robot with Compliant Ankle Joints,
Procs. of the 1997 IEEE International Conference on Intelligent Robots and
Systems, September 7-11, 1997, World Trade Center Atria, Grenoble, France.
7. Keon Young Yi and Yuan F. Zheng, Biped Locomotion by Reduced Ankle
Power, Proceedings of the 1996 IEEE International Conference on Robotics
and Automation, Minneapolis, Minnesota - April 1996.
8. M. Guihard and P. Gorce, Dynamics control of bipeds using ankle and hip
strategies, Procs. of the 2002 I E E E International Conference on Intelligent
Robots and Systems, Lausanne, Switzerland, October 2002.
 WALKER SYSTEM WITH ASSISTANCE DEVICE
FOR STANDING-UP

D. CHUGO*, W. MATSUOKA' and K. TAKASE'

Graduate School of Information Systems, The University of Electro-Communications,
Chofu, Tokyo 182-8585,Japan
*E-mail: chugo, matsuoka, takaseOis.uec.ac.jp
http://www. taka.is.uec.ac.jp/

This paper proposes a walker system with power assistance device for standing
up motion. Our system focuses on family use for aged person who needs nursing
in their daily life. Our key ideas are two topics. The first topic is new assistance
manipulator mechanism with four parallel linkages. Our proposed manipulator
mechanism requires only smaller actuators and realizes rigid structure with
lighter linkages comparing with general manipulator. Thus, we can design our
assistance system compactly with low-cost using our mechanism. The second
topic is the combination of force and position control. According to the patient's
posture during standing up, our control system selects more appropriate control
method from them. We use the reference of standing-up motion which is based
on the typical standing up motion by nursing specialist for realizing the natural
assistance. The performance of our proposed assistance system is verified by
computer simulations and experiments using our prototype.

Keywords: Force assistance; Active Walker; Standing up motion; Force control;

Position control,

1. Introduction
In Japan, the population ratio of senior citizen who is 65 years old or
more exceeds 20[%] at January 2004 and rapid aging in Japanese society
will advance in the future.' In aging society, many elderly people cannot
perform normal daily household, work related and recreational activities
because of decrease in force generating capacity of their body. Today, the
23.5[%] of elderly person who does not stay at the hospital cannot perform
daily life without nursing by other people.2 For their independent life, they
need assistance system which enable them to perform daily life easily in
their home even if their physical strength reduces.
Standing up motion is the most serious and important operation in daily

475
476

life for elderly person who doesn’t have enough physical ~ t r e n g t h . In ~?~
typical bad case, elderly person who doesn’t have enough physical strength
will cannot operate standing up motion and will falls into the wheelchair
life or bedridden life. Furthermore, if once elderly person falls into such life,
the decrease of physical strength will be promoted because he will not use
his own physical ~ t r e n g t h .Therefore,
~ force assistance system is required
t o use part of the remaining strength of the patient for standing-up motion
in order not t o reduce their muscular strength.
In previous works, many researchers developed power assistance devices
for standing up motion. However, these devices are large scale and it is
difficult for family US^.^,^ Furthermore, these devices assist all necessary
power for standing up and they do not discuss the using the remaining
physical strength of patients. Therefore, there is a risk of promoting the
decrease of their physical strength.
In this paper, we develop a walker system with power assistance device
for standing up motion. Our system is based on a walker which is popular
device for aged person in normal daily life and realizes the standing up
motion using the support pad which is actuated by the manipulator with
three degrees of freedom. For using the remaining physical strength, our
system uses the motion pattern which is based on the typical standing up
motion by nursing specialist as control reference.8
Our key ideas are two topics. The first is new assistance manipulator
mechanism with four parallel linkages which enables the system t o be rigid
and compact. The second is the combination of force and position control.
Using our control scheme, the patients can stand up with fewer loads and
can use their own remaining physical strength during the motion. We ver-
ify the performance of our proposed assistance system through computer
simulations and experiments using our prototype.

2. System Configuration
2.1. Assistance Mechanism
Fig.1 shows our proposed assistance system. The system consists of support
pad with three degrees of freedom and the walker system. (Fig.l(a)) The
support pad are actuated by our new assistance manipulator mechanism
with four parallel linkages.
Fig.l(b) shows the frame-kinematic model of our assistance manipu-
lator. The position of the support pad (2DOF, red lines in Fig.l(a)) is
coordinated by Actuator 1 and Actuator 2 which are equipped on 0 point.
 477

Actuator 1 drives Link1 (a)and Actuator 2 drives Link 2 (p). Using four
parallel linkages mechanism, two actuators can generate the position of sup-
port pad. The inclination of the support pad (lDOF, blue lines in Fig.l(a))
is coordinated by Actuator 3 which is equipped on P point.
The advantages of our proposed mechanism are two topics. The first is
that two main actuators (Actuator 1 and 2) are required to keep only own
weight. In general manipulator, the actuator of lower part supports not only
weight of linkages but also actuators of upper part. Therefore, the actuators
of lower part are required high output traction and tend to be heavy. On
the other hand, using our mechanism, main actuators are mounted on the
walker body (0point) and they are required to keep only weight of linkages.
As the result, we can use smaller actuators for our assistance manipulator.
The second is parallel linkage mechanism. The parallel linkage mech-
anism is strong for a twist load and we can use lighter linkages for our
proposed manipulator comparing with the general one. Using our proposed
mechanism, we can use smaller actuators and lighter linkages, and our sys-
tem realizes compact design with low cost.

(a) Prototype system (b) Parallel Linkage Mechanism

Fig. 1. Our assistance system

2.2. ~ i ~ e ~ a t i c s
In this section, we derive the inverse kinematics of our proposed linkage
mechanism. Using our proposed mechanism, the position of P point ( z pyp) ,
is derived as follows;
The first, we set angular values and length of linkages as in Fig.l(b).
478

el = -7r2 + p , e2 = -7r2 + a - p (1)

11 = L2, 12 = L3 - L1 (2)
Now, we consider the geometric relationships among the position of P
point and these angular values, we can derive (3) and (4). From (3) and
(41, e2 is (5)

x p = l1 cos el+ i2 cos (el + e,) (3)

yP = 11 sin 01 + sin (el + e,)
12 (4)
x; + y; - 1:: - 1;
82= arccos (5)
21112
We set kl and k2 as (6), (3) and (4) are expressed as (7) and (8).

kl = 11 + 12 cos 02, k2 = l 2 sin O2 (6)

x p = kl cos O1 - k2 sin O1 (7)
yp = kl sin O1 + k2 cos el (8)
Furthermore, we set r and y as (9), kl and k2 are expressed as (10).
Using (lo), (7) and (8) are expressed as (11) and (12).

k1 = r cos y, k2 = r s i n y

xP = r c o s ( y + e l )

yP = rsin (y + 01)
From (11) and (12), we can derive (13).

tan(y + 0,) = -
YP
(13)
XP
Thus, 81 is (14). Using ( 5 ) and (14), we can control our manipulator.

el = arctan (E) - arctan (i1 + z2

12 sin 02
)
cos e2 (14)
 479

3. Motion Control
3.1. ~ o t i o n
by Nursing Spe&~alist
In previous study, a lot of standing up motions for assistance are proposed.
Kamiyag proposed the standing up motion which uses remaining physical
strength of the patients maximum based on her experience as nursing spe-
cialist. In our previous work, we analyze this standing up motion and find
that Kamiya scheme is effective to enable standing up motion with smaller
load.8
We assume the standing up motion is symmetrical and we discuss the
motion as movement of the linkages model on 2D plane.'' We measure the
angular values among the linkages, which reflect the relationship of body
segments. The angular value is derived using the body landmark as shown
in Fig.2(a).
Fkom measuring results, the trunk needs to incline to forward direction
during lifting up from chair as shown in Fig.2(b). Y-axis shows the angu-
lar value (Pelvis and trunk, knee, ankle) and X-axis shows the movement
pattern (S)ll which means the ratio of standing up operation as (15).

where t s is required time to the standing up operation and t is present time.

180
160
140
.., 120
5 100
'u

1
p 80 0
a 60
40 I-"-
I
20
0 I I
0 25 50 75 100
Movement Pattern (%)

(a) Coordination (b) Angular Value

$1, $2 and $3 show the angles of pelvis/trunk, knee and ankle, respectively.

Fig. 2. Standing-up motion with Kamiya scheme

480

3.2. Derivation of Control Reference

In this section, we derive the control reference of our assistance system
which can realize the standing up motion proposed by Kamiya using a
computer simulation. Fig.S(a) shows the simulation setup. The parameters
are chosen from a standard body data of adult rnalel2.l3 We use the Work-
ing Model 2D as a physical simulator and MATLAB as a controller. Both
applications are linked by Dynamic Data Exchange function.

(a) Simulation Setup (b) Position Reference (c) Angle Reference

Fig. 3. Derived control references

F’rom these simulation results, Fig.S(b) shows position references of the

support pad and Fig.S(c) shows reference angle. The coordination is defined
as shown in Fig.S(a). In Fig.3(b), the start point is lower and the end point
is upper. Using these tracks as the position control reference, our assistance
system can realize the standing up motion which Kamiya proposes.

3.3. Force Control

For realizing the motion of Kamiya scheme using the remaining physical
strength of patients, we propose new control scheme as (16). Proposed con-
trol scheme combines dumping control and position control. The dumping
control is suitable for the control of the objects with ~ 0 n t a c t . l ~

v.t -
-
vref
2
-B ( F - Fo)- K (xi - xyf) ... (if F > Fo)
(16)
vi = vyf - K (xi -xi * * - (if F 5 Fo)
where F is the applied force on the support pad as shown in Fig.S(a) Fo
is the threshold which selects force or position control. vyf is the veloc-
ity reference and x r f is the position reference from track references in
 481

Fig.3(b) and (c). vi is actucal control reference. i is identification number

of actuators. (i = 1,... ,3) B and K are constants.
Using this control scheme, when the patient does not use own physical
strength, the applied force on the support pad will be large and system
will assist him more slowly. Therefore, the patient can use own physical
strength easily.

4. Experiment
4.1. Computer S~mulat~on
We verify the performance of our control scheme by the computer simula-
tion. In this experiment, the human model stands up with Kamiya motion
and our assistance system assists him using our proposed control scheme.
Furthermore, we compare the result by our proposed scheme with the result
using only the position control reference.
We use the control references as shown in Fig.3(b) and (c) which are
derived from standing up motion with Kamiya scheme in previous section.
The simulation parameters are chosen from Table 1. The coordination is
defined in Fig.3(a). Furthermore, we set E;b=l5[N] as a threshold which is
derived experimentally.
Fig.4 shows the simulation results. Fig.4 is standing up motion using
our proposed assistance control. Allows in Fig.4 show the applied assistance
force to the patient. The support pad applies the force vertically to the
chest and the hand. Using our proposed control scheme, we verified that
our assistance system realizes the natural standing up motion.

Fig. 4. Simulation results

Table 2 shows the maximum output and output power of each joint.
Fkom these results, our proposed control scheme reduces maximum out-
put into 0.5[Nrn/kg] comparing with the result by the position control. In
482

general, if the applied load to each joint is heavier than 0.5[Nm/kg], it is

difficult to stand up for the elderly person. Therefore, the patient can adapt
this motion using his remaining physical strength.
Furthermore, with our proposed control scheme, the workload of each
joint is maintained comparing with the result by the position control. Dur-
ing standing up motion, the patient is required to use his 92[%] of physi-
cal strength comparing with physical strength without the force assistance
control. This means our assistance system can use part of his remaining
strength in order not to reduce muscular strength. Therefore, we can verify
that the system selects more appropriate control method and our proposed
control scheme is effective.

Table 1. Simulation results

Pelvis/Trunk I Knee I Ankle
I

PM y Position Conti

Proposed Control
0.48
29.0
0.49
26.9
I 0.59 I 0.38
I
39.3
0.50
36.3
31.0
0.38
28.8

4.2. Experiment
Here, we verify the performance of our prototype system by the experiment.
In this experiment, our prototype system assists the patient with control
references shown in Fig.S(b) and (c). As the result of the experiment, our
system can assist the patient as shown in Fig.5. The height of the patient
is 170[cm] and the system lifts him at 30[sec].

Fig. 5. Experimental results

Fig.6 shows the tracks of the position of the patient’s waist, knee and
ankle joint and their control references. From Fig.6, both tracks are al-
 483

0 20 60 60 80 I00
I Moverntn Pattern (t)

(c) Ankle

Fig. 6. Angular value of each joint

most same line and this means our assistance system realizes the natural
standing-up motion by nursing specialist.

5. Conclusion
In this paper, we develop the novel assistance system for the standing up
motion. Our system focuses on family use and system can assist the elderly
person using part of their remaining strength, in order not to reduce mus-
cular strength. In order to fulfill this condition, we propose new assistance
manipulator mechanism with parallel linkages. Our developed mechanism
enables the assistance system to be rigid and compact with low costs. Fur-
thermore, we design the novel control scheme which combines the dumping
and the position control. According to the posture of the patient during
standing up motion, our control system can select more appropriate con-
trol method from them. Our assistance system realizes the natural standing
up motion by nursing specialist and it is effective to assist the aged person
to stand up without reducing their muscular strength.
484

In our future work, we will develop t h e active walker function, which

enables to assist the patients during standing, walking and seating motion
continuously.

Acknowledgments
This study is supported by France Bed Medical Home Care Research Sub-
sidy Foundation.

References
1. Japan current population estimates as of October 1, 2004 (2004), http://
www.stat.go.jp/english/data/jinsui/2004np/index.htm.
2. Japan annual reports on health and welfare 2001 social security and
national life (2001), http: //www.m h l w , go. jp/toukei/saikin/hw/k-tyosa/
k-tyosaOl/4-3.html.
3. N. B. Alexander, A. B. Schultz and D. N. Warwick, J. of Geometry: MEDI-
CAL SCIENCES 46, M91 (1991).
4. M. A. Hughes and M. L. Schenkman, J. of Rehabilitation Research and De-
velopment 133,409 (1996).
5. M. Hirvensalo, T. Rantanen and E. Heikkinen, J . of the American Geriatric
Society 48, 493 (2000).
6. K. Nagai, I. Nakanishi and H. Hanabusa, Assistance of self-transfer of patients
using a power-assisting device, in Proc. IEEE Int. Conf. on Robotics and
Automation (ICRA ’03), (Taipei, Taiwan, 2003).
7. A. Funakubo, H. Tanishiro and Y. Fukui, J . of the Society of Instrument and
Control Engineers 40, 391 (2001).
8. D. Chugo, E. Okada, K. Kawabata, H. Kaetsu, H. Asama, N. Miyake
and K. Kosuge, Force assistance control for standing-up motion, in Proc.
IEEE/RAS-EMBS Int. Conf. on Biomedical Robotics and Biomechatoronics
(BioRob’Oti), (Pisa, Italy, 2006).
9. K. Kamiya, Development and evaluation of life support technology in nursing,
in Proc. 7th RACE Symp., Research into Intelligent Artifacts for the Center
of Engineering, (Chiba, Japan, 2005).
10. S. Nuzik, R. Lamb, A. Vansant and S. Hirt, J . of Physical Therapy 66, 1708
(1986).
11. P. 0. R. R. W. M. M. Schenkman, R. A. Berger and W. A. Hodge, J . of
Physical Therapy 70,638 (1990).
12. Human body properties database (digital human research center of aist)
(2006), http: //www. stat .go.jp/english/data/jinsui/2004np/index.htm.
13. K.Omori, Y. Yamazaki, H. Yokoyama, U. Aoki, M. Kasahara and K. Hiraki,
Sogo Rehabilitation 30, 167 (2001).
14. T. Sugihara, K. Kawabata, H. Kaetsu, H. Asama, K. Kosuge and T. Mishima,
Development of a reasonable force sensor for a standing-up and sitting motion
support system, in Proc. of Robotics and Mechatronics Conf. 2004, (Nagoya,
Japan, 2004).
Advances in Humanoid Soccer Robots
This page intentionally left blank
 A Distributed Embedded Control Architecture for Humanoid
Soccer Robots.
Carlos Antonio Acosta Calderon, Changjiu Zhou, Pik Kong Yue,
Mike Wong, and Mohan Rajesh Elara
Advanced Robotics and Intelligent Control Centre,
School of Electrical and Electronic Engineering,
Singapore Polytechnic, 139651, Singapore
Email: { CarlosAcosta, ZhouC) @sp.edu.sg
www.robo-erectus.org

This paper presents an embedded control architecture developed for a soccer-

playing humanoid robot, namely Robo-Erectus Junior. Robc-Erectus is a
project developed in the Advanced Robotics and Intelligent Control Centre
of Singapore Polytechnic that will participate in the KidSize category in the
Humanoid League of RoboCup 2007. The Robo-Erectus project has reached
a new stage with this latest version of the robot. The new mechanical, elec-
tronic design, and embedded control architecture are described in this paper.
The limitations of size force us to implement this architecture in an embedded
network. The features of this new version are described in detail focusing on
the modules of gait generation, vision, behaviour control, and communication.

Keywords: Behaviour Control, Humanoid Robot, Soccer Robots

1. Introduction
The Robo-Erectus project is developed in the Advanced Robotics and In-
telligent Control Centre of Singapore Polytechnic. Robo-Erectus is one of
the foremost leading soccer-playing humanoid robots in the RoboCup Hu-
manoid League.' Robo-Erectus has collected several awards since its first
participation in the Humanoid League of RoboCup in 2002. Some of the
awards include the 2nd place in the Humanoid Walk competition at the
RoboCup 2002, the lStplace in the Humanoid Free Performance compe-
tition at the RoboCup 2003. In 2004, Robo-Erectus won the Znd place in
Humanoid Walk, Penalty Kick, and Free Performance.
The competitive level of the humanoid teams in RoboCup has increased
and the challenges in the competition are becoming tougher every year.
For the Robo-Erectus project to keep to as one of the leading teams is

487
488

imperative to have a robust and reliable control system that allows the
robot to exploit its full capabilities. This paper provides the inside of the
architecture for the Robo-Erectus Junior which controls not just a robot
but display the emergent behaviour of a team.
The development of Robo-Erectus has gone through many stages either
in the design of its mechanical structure, electronic control system and
gait movement control. The latest version of Robo-Erectus, namely Robo-
Erectus Junior, has been designed to cope with the complexity of a soccer
game. Robo-Erectus is able to perceive different colours and to track them.
The robot also contains a dedicated processor used to control its behaviour,
wireless communication with the control PC and teammates, and a sub-
system to control sensors and actuators. For more detailed information
about the Robo-Erectus humanoid soccer robots, please refer to the team’s
website www.robo-erectus.org.
This paper describes the current state of the project developed for the
2007 RoboCup competition. The rest of the paper is organized as follows.
In the next Section, details of the mechanical and electronics design are pre-
sented. Section 3 describes the software developed: the image processing,
the hierarchy of the control system for the robot behaviour, and the infras-
tructure needed to support a team of soccer playing robots, respectively.
Finally, some concluding remarks are presented in Section 4.

2. Hardware Design
The Robo-Erectus project aims to develop a low-cost fully-autonomous
humanoid platform that could be used for competition, education, and
research purposes.2 The previous generations of humanoid soccer robots,
namely RE401, RE4011, RE40111 and RE70; have provided us with a robust
knowledge about the hardware and software design, This Section provides
an inside to the latest development namely Robo-Erectus Junior. Appendix
A presents a compendium of the specifications of the Robo-Erectus Junior.

2.1. Mechanical Design

Figure 1(a) shows the design of the humanoid robot Robo-Erectus Junior.
The skeleton of the robot is constructed with aluminum braces. The head
and arms of the robot are made of plastic. Despite its simplicity, the me-
chanical design of the robot is robust and lighter weight in comparison with
its predecessors. Its human-like body has a height of 50cm and weight of
just 3.2kg, including batteries.
 489

Fig. 1. Robo-Erectus Junior, the latest generation of the family Robo-Erectus.

Robo-Erectus Junior has a total of 24 degrees of freedom. Further details

of the body parts and their associated degrees of freedom can be found in
Appendix A. Each degree of freedom uses as actuator a D ~ n a m ~DX-117
~el
Digital Servomotor. These servomotors have a typical torque of 28.89kg e r n
and a speed of 0.172sec/6O0. Each smart actuator has a micro-controller
in charge of receiving commands and monitoring the performance of the
actual motor. An RS485 serial network connects all the servomotors to a
host processor which, sends positions and receives the current data (angular
positions, speed, voltage, and temperature) of each actuator.

2.2. ~ l ~ ~ t rDesign
o n i ~
Figure l(b) shows the network of three processors and the main devices
that controls the Robo-Erectus. The use of processors for particular tasks
improves the performance of the system. Each processor is dedicated to
different tasks:
(1) The main processor coordinates all behaviors of the robot, collects the
sensor information and sends the data t o the motors, and communicates
by WIFI with the other robots and the main PC.
(2) The vision processor processes the image obtained from the USB cam-
era. This processor is connected with the main processor by RS232.
490

(3) The Sensor/Actuator dual micro-controller receives the motor com-

mands from the main processor. These commands are validated and
finally sent to the servomotors by RS485. The motor feedback is col-
lected and sent back to the main processor every 16.6ms. This system
is also responsible to collect the sensors' values to be sent to the main
processor.
Robo-Erectus has four main sensors: a camera to capture images, a
tilt sensor to recognize whether it is standing or falling down, a compass
to detect its own orientation, and finally a couple of ultra-sonic sensors to
measure distance from front and rear objects. As mentioned earlier, the ser-
vomotors send back the feedback data i.e. angular positions, speed, voltage,
and temperature. To communicate with its teammates, Robo-Erectus uses
a wireless network. The WIFI interface is connected to the main processor.
The connection of the main building blocks can be seen in Fig. l(b).
Finally, Robo-Erectus is powered by two high-current Lithiumpolymer
rechargeable batteries, which are located in each foot. Each battery cell
has a weight of only llOg providing 12V which is able to operate about 20
minutes.

3. Control Architecture
The control architecture for each robot is distributed in a network of three
processors. The main processor runs Linux as operative system. Due to
the limitations of the system the footprint of the Embedded Linux is very
small, but yet powerful to permit to take all the advantages of this operative
system, such as threading, networking, connectivity, and so forth. The team
behaviour in a game emerges form interaction of the robots and the role
selection. The functionality of Robo-Erectus is divided in several software
modules, which are described in detail below.

3.1. Motion Control

Our team has studied different approaches to motion control, includ-
ing kinematics, dynamics, fuzzy logic, neural networks and genetic algo-
r i t h m ~ . ~How
- ~ to generate a dynamically stable gait for the humanoid
soccer robots with consideration of various constraints is still an important
research topic in this area.
Our latest approach employs the Estimation of Distribution Algorithm
EDA for gait ~ptimization.~!'This approach speeds up the searching in a
highly dimensional coupling space constructed by the permutation of the
 491

optimization parameters to establish a periodic orbit in biped locomotion.

Based on the maximum entropy principle, we also developed a Factorized
D ~ s t r i b u t ~ oAnl g o r i t ~FDA
~ based gait optimization method to better un-
derstand how information are transferred between these parameters so that
we may progress toward better understanding human locomotion and ex-
tend the results to design of humanoid robot^.^^' We used the EDA and
FDA have been successfully used to generate biped gaits that satisfy a cri-
terion. The gaits have been efficiently used to drive the humanoid robot
Robo-erect us Junior.

Fig. 2. The suit of in-house development tools.

These gaits and other behaviours are programmed by using a IDE soft-
ware, which controls from a single servo position to a full complex be-
haviour. This IDE allows that any changes to the gait can be implemented
and tested in short turnaround time (See Fig. 2). In addition, the data of
each joint can be monitored and analyzed in real time. In order to be able
to design behaviors without access to the real hardware, we implemented
a physics-based simulator for the robots. This simulation is based on the
Open Dynamics Engine.lo

3.2. Image Processing

The vision system provides the main source of information about the en-
vironment for the robot. The system consists of a 70" camera mounted on
492

a pan-tilt system that allows the robot to scan 240' wide. The computa-
tion of the vision system occurs in a dedicated embedded system, only the
output is transmitted to the main system.
The main process on the image is colour s e g ~ e n t a t ~ obased
n on the HSI
colour space. Each image's line is scanned pixel by pixel. During the scan,
the pixels are classified by color. A look-up-table is employed to classify the
pixels by colours described by an HSI range for each colour.

Fig. 3. The image processing

In a multistage process we discard insignificant colored pixels and detect

colored objects (See Fig. 3). Recognition algorithms for the most important
features (i. e, lines, goals, corners, robots, and the ball) axe implemented
avoiding a wrong identification. In addition, previous observations of these
objects provide confidence for their identification.
Finally the system estimates the position of each object in an egocen-
tric frame (distance from the robot and its orientation). This egocentric
map is merged along with sensory data and teammate information into a
allocentric map, which provides a robust world representation.
The gaze control uses a fovea as main tracking area. When the object
of interest is found inside the fovea the pan and tilt motor will not move.
On the contrary, if the object of interest is outside the fovea, but still in
the image, the pan and tilt motor will move to compensate the distance
from the center of the fovea. If the object is not found on the image, it is
reported and the motor will not move.
A system helps to turn the vision system when the robot is used in
different light conditions. This system uses a remote connection via wireless
network to turn the thresholds that determine a particular colour.
 493

3.3. Behaviour Control

The module provides the control and strategy for the autonomous mode
of the robot when playing. A framework of hierarchical reactive behaviours
is the core of this control module.ll This structure restricts interactions
between the system variables and thus reduces the complexity. The control
of the behaviours happens in three layers: skill, reactive, and planning layer
(See Fig. 4(a)).

Event t 4 Task

(4 (b)

Fig. 4. The behaviour control has been implemented as a hierarchical reactive frame-
work, which is organized in three layers.

The skill layer controls the servos, monitors targets, actual positions,
and motor duties. The skill layer receives actions from the reactive layer and
convert them into motor commands. After performing the motor commands
a feedback is sent back to the reactive layer.
The reactive layer implements the robot behaviours like walking, kick-
ing, getting-up, and so forth. This layer determines the parameters for the
behaviour and these parameters can be adapted on time. This makes pos-
sible to correct deviations in the actual data and t o account for changes
in the environment by using the sensor feedback. Each of these behaviours
consists of several actions, which are sent to the skill layer. The selection
of the behaviours depends on the desire task of the planning layer. The
494

behaviours in this layer are implemented as a subsumption architecture

to enable the robot to satisfy the task while it can navigate safely in the
environment.
The planning layer used the behaviours of the reactive layer to imple-
ment some soccer skills like approaching the ball, dribbling, attacking and
defensive behaviors. The planning layer guide the robot to coordinate its
efforts with the teammates to score goals and defend their goal. The be-
haviours at the planning layer are abstract goals. These abstract goals are
passed to the reactive layer to be sent to the actuators.
Each robot has the same hierarchical reactive framework. The emergent
team behaviour is the product of communication and cooperation between
players. In the game each robot has a particular role. However, the robots
can dynamically switch theirs roles. The robots change theirs roles based
on a particular strategy.
Each role is defined in the planning layer as a state-machine with tran-
sitions triggered by a combination of sensorial information, strategy, and
messages received from the teammates. The implemented roles are goalie,
defender, and attacker; Fig. 4(b) shows the state-machine for the goalie
role. Each robot takes decisions based on its role, perceptions, and the in-
formation it receives from teammates. The robot broadcasts its state to the
teammates whom use this information to decide on role.

3.4. Communication
Robo-Erectus is equipped with wireless network adapters. The robots com-
municate with each other to negotiate roles and to share perceptions. The
wireless communication is also used to transmit information to an external
computer for recording and visualization. A monitor system was developed
to follow the state of the robots during a game. This monitor system also
broadcasts to the players information about the game, e.g. kickoff, finish,
penalty, free-kick. The information received from each robot can be recorded
for analysis purposes.

4. Conclusion
In this paper, we introduced the state-of-art of the Robo-Erectus project.
Robo-Erectus Junior is an autonomous humanoid robot with a network
of three embedded systems, 24 degrees of freedom, and several kinds of
sensors that serves as a platform of education, edutainment, and research
issues. The latest version of the Robo-Erectus holds several advantages
 495

in contrast with the previous generations, e.g. faster, robust control, gait
improvement, vision. The new features prepare the Robo-Erectus for the
2007 RoboCup competition. New development tools were conceived from
the gained experience of the previous versions of the Robo-Erectus. In ad-
dition, the improvement in the robot platform allows a more robust and
efficient performance of the robot in the autonomous mode. Research with
this platform has lead to develop a new approach to better control of hu-
manoid robots in particular for RoboCup.

Appendix A. Mechanical Specifications of Robo-Erectus Junior

I

Dimensions Speed
Weight
Height Width Depth Walking Running
3.2kg 480mm 270mm 150mm 2mts/min __

Joint Roll Pitch Yaw

Head J J
Body J J
Shoulder J J J
I I I

Bir, J J I J I

Sonar 1 Distance range 5cm-250cm.

4. Specifications of the boards of the kbo-Erectus Junior.
Features Main Processor Vision Sensor/Actuator
Processor Intel ARM XScale Intel ARM XScale Dual PIC18F8720
Speed 4OOMhz 400Mhz 25Mhz
Memory 16MB 32MB 8KB
Storage 16MB 16MB 256KB
Interface RS232, WIFI RS232, USB RS232, RS485
496

Acknowledgements
The authors would like t o thank staff and students at the Advanced
Robotics and Intelligent Control Centre (ARICC) and higher management
of Singapore Polytechnic for their support in t h e development of our hu-
manoid robots. I n particular t o Chin Keong Ang, Weiming Yuan, Edric Lee
Lai Fatt, Guohua Yu, Weijie Ye, Yang Bo and Aung Aung Kyaw for their
contributions t o this project.

References
1. H. Kitano and H. Asada, The RoboCup humanoid challenge as the millen-
nium challenge for advanced robotics, Advanced Robotics 13, (2000).
2. C. Zhou and P. Yue, Robc-Erectus: A Low Cost Autonomous Humanoid
Soccer Robot, Advanced Robotics 18, (2004).
3. Z. Tang, C. Zhou and Z. Sun, Gait synthesizing for humanoid penalty kicking,
in Proc. 3rd International DCDIS Conference o n Engineering Applications
and Computational Algorithms, (Ontario, Canada, 2003).
4. C. Zhou and Q. Meng, Dynamic balance of a biped robot using fuzzy rein-
forcement learning agents, Fuzzy Sets and Systems 134, (2003).
5 . L. Hu, C. Zhou and Z. Sun, Biped gait optimization using estimation of dis-
tribution algorithm, in Proc. of IEEE-RAS Int. Conf. on Humanoid Robots,
2005.
6. L. Hu, C. Zhou and Z. Sun, Biped gait optimization using spline function
based probability model, in Proc. of IEEE Int. Conf. on Robotics and Au-
tomation, (Orlando, USA, 2006).
7. L. Hu, C. Zhou and Z. Sun, Estimating probability distribution with Q-
learning for biped gait generation and optimization, in Proc. of IEEE Int.
Conf. on Intelligent Robots and Systems, 2006.
8. L. Hu, C. Zhou and Z. Sun, Hybrid Estimation of Distribution Algorithm
with Application to Biped Gait Generation, Information Sciences (2006).
9. C . Zhou, L. Hu, C. Acosta and P. Yue, Humanoid soccer gait generation and
optimization using probability distribution mode, in Workshop on Humanoid
Soccer Robots, Proc. on the IEEE International Conference on Humanoid
Robots, (Genoa, Italy, 2006).
10. Open dynamics engine, www.ode.org url.
11. S. Behnke, B. F'rotschl, R. Rojas, P. Ackers, W. Lindstrot, M. de Melo,
A. Schebesch, M. Simon, M. Sprengel and 0. Tenchio, Using hierarchical
dynamical systems to control reactive behavior, in RoboCup, 1999.
 Design of a Humanoid Soccer Robot: Wukong

Qing Tang, Rong Xiong, Jian Chu, Xinfeng Du

State Key Lab. of Industrial Control Technology
Zhejiang University
Zheda Road No.38, Hangzhou 310027
Zhejiang Province, P. R . China
tangq@iipc.zju.edu. c n
http://www. nlict.zju. edu. cn/humanoid/index. html

This paper presents a complete robot system for humanoid soccer robot,
Wukong. 22 DOFs and 2 camerm are designed for the robot to take part in the
soccer game. Specific hardware architecture is introduced. Robot locomotion,
robot vision and robot decision are all achieved on the robot.

Keywords: Humanoid Soccer Robot; Biped Locomotion; Robot Vision; Robot

World Model; Wukong

1. Introduction
Recent years, the humanoid robot research has drawn hundreds of re-
searchers’ interests and some of their robots have been acting as the re-
ceptionist or can deliver drinks for people. The most exiting humanoid
robots include the Asimo made by Honda,l Wabian developed in Waseda
University12HRP-3P from AIST13KIST’S KHR-3* and some other research
groups’ robots. They are leading the research of humanoid robots in robot
design and locomotion.
At the same time, motivated by the rapid development of the artificial
intelligence research and robotics, the RoboCup Federation organizes inter-
national robotic soccer competitions since 1997. The long-term goal of the
RoboCup Federation is to develop a team of humanoid soccer robots that
wins against the FIFA world C h a m p i ~ nAfter
. ~ 10 years’ research, the most
attractive match has focused on the humanoid soccer league which started
from 2002. This paper present the design of a robot which is capable to
attend the humanoid soccer competition. The rest of the paper contains
the summarized descriptions of each components of the robot. Mechani-
cal specifications, hardware specifications, and software specifications are

497
498

described in Section 2, 3, and 4 respectively.

2. Mechanical Design of the Robot

In order t o design a humanoid soccer robot t o play soccer, we should enable
our robot Wukong capable of accomplishing the following movements in
physical characters.

0 walk straight
0 turn around
0 Kick the ball
0 catch the ball
0 get up from the ground

Table 1 shows the specification of our robot Wukong. Wukong is driven

by 22 servo motors: 6 per leg, 3 in each arm, 2 in the trunk and 2 in the
head. The six leg-servos allow for flexible leg movements. Three orthogonal
servos constitute the 3-DOF hip joint. Two orthogonal servos form the 2-
DOF ankle joint. One servo drives the knee joint. The pitch and role trunk
joints are specifically designed for humanoid soccer robot. It is much easier
for the robot to get up having the pitch DOF. And the role joint enable the
robot to walk fast. The head is manipulated by 2 motors, so the camera
can turn in quite wider direction as human’s head. Figure l ( a ) shows the
distribution of the DOFs of Wukong.

Dof

Networked smart motor DX-117 is selected as the actuator, because

it offers a high torque with less weight. The motor is controlled with an
in-built microchip. The communication is through RS485 protocol with a
transfer speed up to 1MHz. We connect the motors one by one with wires
and control the motors by their unique ID. With this type of motor we
designed our humanoid soccer robot Wukong which can be seen in fig l(b).
 499

(a) The DOF distribution (b) Our robot Wukong

Fig. 1: Robot’s Mechanics

3. Hardware Architecture
To motivate the 22 pieces of motors move simultaneously, we choose AT-
Mega128 as its microcontroller. ATMegal28 is an 8-bit processor worked
at 16MHz with 128K Bytes of In-System Reprogrammable Flash and 4K
bytes EEPROM. The Baud Rate of the serial port can run up to 1MHz.
The main controller sets the motors’ position one by one and broadcast an
instruction to motivate all the motors together every cycle during robot’s
locomotion.
Vision is provided by two stereo analog cameras using IEEE 1394 inter-
face for output and power. The cameras are 640x480 progressive scan CCD
with a standard miniature lenses. The frame rate is 32Hz for 640x480 with
automatic control of exposure, gain and black level. The image processor
DSP DM642 is used to take image data from the camera. Image threshold
segmentation, shape description and target identification would be done on
this processor. The reason we choose DSP as our main processor is that
DSP has a good capability of processing image data rapidly compared to
500

the other processors such as ARM and PDA. Some basic algorithms are im-
plemented by hardware. On the other hand, DSP DM642 can connected to
the Ethernet based on IEEE 802.11b protocol, which makes it very conve-
nient to communicate between robot teammates and easy for the humanoid
soccer robots to receive the referees’ signals. The overall hardware archi-
tecture of Wukong can be seen in fig. 2.

Fig. 2: Hardware Architecture

4. Software Architecture
What makes a man a good soccer player, not only how strong, quick and
agile he is that matters, but also he must distinguish and locate the target
clearly and accurately as well as make the right decision according t o the
situation of the playground and teammates. So software architecture should
be divided into three parts: gait planning, image processing and decision
making.

4.1. Gait Planning

To generate the agile movement for the robot, ZMP(Zero Moment Point)
C r i t e r i ~ nproposed
~>~ by Vukobratovic is considered. The robot is simplified
by link model and its kinetic and kinematic movement are described using
 501

conventional Denavit-Hartenberg coordinate. Under these hypothesis, the

robot’s gait is generated in the following steps:

0 Preset the lower limbs’ trajectory

0 Preset the ZMP’s trajectory
0 Optimize the hip’s position towards the minimum ZMP domain
based on the robot’s structural parameters and the preset limbs’
trajectory.
0 Plan the upper limbs’ trajectory t o compensate the robot’s dy-
namic balance.
Change the trajectory data from Cartesian coordinate into joints
coordinate.

After getting the trajectory of the left ankle ( x a L ( t ) , y a ~ ( t ) )the

’ trajec-
tory of the right ankle ( x , ~ ( t )Y, a R ( t ) ) ,and the movement of the ZMP point
~ , , ~ - ~ ~ f the
( t )hip
, trajectory (x(t)’y ( t ) ) is obtained considered the geo-
metrical limitation and the optimization of the ZMP criterion. The problem
could be described as follows:
2
min llf(x, -)x z M P - - r e f I I
X,L, Y ~ LX, ~ R ’ Y ~ R

{ (XI Y)
21
= g ( X a L l Y ~ LXI a R , Y a R )
5 x 5 X2’Yl I Y 5 YZ
Function f returns the ZMP position of the robot under given hip point
and function g constitutes the geometrical limitation.
Computation of the hip’s position is an optimization problem of func-
tional extreme value. If each generated hip position would lead the actual
ZMP approaching t o the objective ZMP, we could admit that the generated
hip trajectory would also lead the actual ZMP trajectory approaching to
objective ZMP trajectory. So the optimization problem of functional ex-
treme value could be converted into objective function optimization prob-
lem which is easier in computing. The method could be expressed in the
following way.

I f(?(tl))

f( ?( tn))
+

+
zzrnp--ref(tl)
f(z(t2))+ X z r n p - r e f ( t 2 )

xzmp--ref(tn)
* f(z(t)) -+

t l , t ~. . ,. t, stands for the sample time intervals of the planning gait.

The hip trajectory results in a stable gait. Robot achieved the basic
gait of walking, turning, kicking the ball, catching the ball and standing
xzrnp--ref(t)
502

up using the method. To verify the robot motion, we choose Adams which
is a dynamical mechanical simulation software as our simulator. First we
create a model based on our robot’s structure. Then Adams takes the joints
angle’s trajectory as the input of the model, considers the robot’s inertial
effect and the counterforce of the ground, and outputs the stability of the
robot. It is convenient and useful to have the gait tested using this software,
as well as it avoids some potential damage to the robot.

4.2. Image Processing

Cameras worked as the robot’s eyes are the main feedback of the humanoid
soccer robot. It is an essential capability for the robot to estimate its pose
on the field robustly and accurately. The main difficulties are that the robot
has a constrained field of view, that the role and pitch angle of the camera
changes continuously and can only be roughly estimated, and that the colors
of the landmarks are usually affected by the lighting and the pose of the
camera. Current work has been focused on solving the last problem with
self-adaptive threshold segmentation. In this method, the threshold values
are updated every sample time, processor records the history thresholds and
associated weights for each target. The current thresholds are pre-calculated
based on the historical records as follows:

ThresholdcuTTent
=
C (HistoryThreshold [i] * Weight [i])
C Weight [i]
With these pre-thresholds and shape description, targets in the color im-
age are identified. A new threshold value and associated weight of targets
are calculated according to the reliability of identification. The weight of
the threshold decrease autonomously and both of them would be discarded
after fixed times. Much better results have been attained using the adap-
tive threshold for segmentation. Then we can specify the distance of the
target according the position and orientation of the cameras. More specific
research has been carrying on.

4.3. Decision Making

Decision making is made based on what the robot learned about the envi-
ronment. It covers updating the robot’s world model and path planning.
The robot’s world model contains all the targets’ position and the robot
itself’s pose. It is constructed according to the fusion of perception and
locomotion of the robot. The multi data fusion change the data from polar
 503

coordinate into Cartesian coordinate. Because the robot is not equipped

with omni-camera, limited vision is obtained, the most accurate and trusted
landmarks are selected based on the robot’s world model and the image
process results. Using this landmarks, robot could relocate itself and other
objectives and update its world model every frame of image.
Attack phrase and defensive phrase are defined in the world model. Each
robot’s moving path is calculated to optimize the shortest time considering
the locomotion. Robots would communicate with each other using wireless
network to balance their strategies.

5. Experiment
Experiment is carried out on Wukong. Maximum walking speed of 7cm/s
is achieved and all the movements listed above are all realized. Fig 3 shows
the identification of the ball on the playground. Fig 4 shows our humanoid
soccer robot Wukong is kicking the ball.

Fig. 3: Identification of a ball on the playground.

A black line indicates the radius of the ball.

Much more videos of our robot could be found on the web site:
504

Fig. 4: Wukong is kicking the ball

http://www.nlict .zju.edu.cn/humanoid.

6. Conclusion
This paper presented the development of the humanoid soccer robot
Wukong. It was proved that Wukong provided with the current mechanical,
hardware and software control system can walk, turn around, kick the ball,
catch the ball, and get up from the ground autonomously and is qualified
as a soccer player for RoboCup 2007.
In the future, we would intend t o provide Wukong with more compli-
cated movements and much better understanding of the world. Further-
more, the analysis and improvement of the hardware and software archi-
tecture will also be continued.

References
Y. Sakagami, R. Watanabe, C. Aoyama, S. Matsunaga, N. Higaki and K. fi-
jimura, The intelligent asirno: system overview and integration, in Proceedings
of 2002 IEEE/RSJ I n t e ~ u t ~ o Conference
~al on r ~ t e l ~ ~ g Robots
ent and Sys-
tem, 2002.
Y . Ogura, H. Aikawa, K. Shirnomura, H. Kondo, A. Morishima, H.-o. Lirn
and A. Takanishi, Development of a new humanoid robot wabian-2, in the
2006 IEEE Internationul Conference on Robotics and A ~ t o m a t ~ o (Orlando,
n,
Florida, 2006).
 505

3. K. Akachi, K. Kaneko, N. Kanehira, S . Ota, G. Miyamori, M. Hirata, S . Kajita

and F. Kanehiro, Development of humanoid robot hrp-3~2005.
4. P. Ill-Woo, K. Jung-Yup, P. Seo-Wook and 0. Jun-Ho, Development of hu-
manoid robot platform khr-2 (kaist humanoid robot-2)2004.
5 . H. Kitano and M. Asada, The RoboCup Humanoid Challenge as the Millen-
nium Challenge f o r Advanced Robotics, Advnaced Robotics 13,723 (2001).
6. M. Vukobratovic and D. Juricic, Contribution to the synthesis of biped gait,
IEEE Tkansactions o n BioMedical Engineering 16,1 (1969).
7. M. Vukobratovic and B. Borovac, Zero-moment point - Thirty five years of its
life, International Journal of Humanoid Robotics Vol. 1, 157 (2004).
 FORMULATION OF DESIRED ZERO MOMENT POINT
TRAJECTORY USING STATISTICAL METHOD

LINGYUN HU, CHANGJIU ZHOU, BI WU, TIANWU YANG

School of Electrical and Electronic Engineering, Singapore Polytechnic
Dover 500 Road, 139651, Singapore
E-mail: ZhouCJ@sp.edu.sg
http://www.ro bo- erectus. org/

This article explains how to construct desired Zero Moment Point (ZMP) re-
gion by statistical method. Gait samples, generated by humanoid soccer robot
named Robc-Erectus Senior, are converted to unified ones with standard step
length. They are evaluated by ZMP margin and energy consumption. The es-
timated values are also used as weights in the calculation of means and covari-
ance. Thereby, ZMP quality in term of dynamical stability and energy efficiency
can be evaluated by covariance. The less the covariance is, the better the ZMP
is. For given covariance, desired ZMP region can be constructed easily and
quickly for further optimization and learning.

Keywords: humanoid robot; desired ZMP;

1. Introduction
Intuitional evaluation on optimum performance of biped walking is max-
imum walking speed and minimum power consumption. To be more spe-
cific,it can be quantified by the following criterions.
(1) Balance Biped walking is supposed to be stable and maintains balance
during the locomotion;
(2) Efficiency Dynamical load of each robot joint should be balanced and
harmonious;
(3) Anthropomorphosis Gaits are not only feasible but also hominine.
Anthropomorphic characteristics] like the flexion of knees should be
considered in gait control methods.
Well balanced gaits with minimum energy cost and hominine characters
are called desired gaits. They have wide application in biped gait generation
and optimization as the reference ones [l,21.

506
 507

Since biped robot tends to fall in nature because of the unilateral and
underactuated foot joint [3], criterion (1) is the basis t o generate realizable
gaits for biped mechanism. Indicators of the degree of dynamical balance
has been profoundly interwind [2, 3, 41. For dynamical stability, the ZMP
relative position with respect to the support footprint is popularly used in
evaluation and control [2].
Research in biological motion pattern has shown that biological infor-
mation, like structural information and dynamic cues, is encoded in biped
gaits [5]. It was reported that, discrimination of biological motion mediated
by gender is supposed to be determined mainly by hip width and shoulder
width [6]. In other words, candidate has personal gaits carrying individual
information for recognition. Like human, biped robots have specified limb
and mass distributions, which make the individual difference in their per-
sonal gaits. Thereby, desired ZMP, whose corresponding biped gait fits the
designated robot best, has personal characteristics.
Fast motion pattern generation techniques that follows the desired tra-
jectory in form of ZMP trajectory have been proposed for different digital
and physical robots [7, 11. Moreover, many works have been devoted to
define the optimal desired gait of biped robot based on minimizing the in-
tegral of ZMP displacement. It always yielded a nonlinear programming
problems, that can be solved by artificial intelligent methods [8] or opti-
mization algorithms [9].
However, those methods were all absence of generation and analysis of
desired ZMP trajectories. Generally, to realize walking with large stability
margin, the middle line in the stable region is chosen as the desired ZMP
trajectory [lo]. It is discontinuous when shifting from double support phase
to single support phase, unless such a shift is constrained to happen only
when ZMP is exactly in the center of the foot about to become the single
support one. Actually, in the absence of noise, any trajectory that stays
within stable region would be equally viable in term of dynamical stability.
In addition, gaits with such kind of ZMP trajectories are infeasible in energy
cost of impulsive control. The interested aim of this paper is to construct
the desired ZMP trajectory region using statistical method.
Besides ZMP margin, the second criterion on energy cost is also consid-
ered in ZMP trajectory evaluation and desired ZMP region construction.
For human walking, a minimum expenditure curve with appropriate com-
binations of speed and step length has been reported [ l l ] . Similar to that,
mechanical energy expenditure can be evaluated by the time integral of the
instantaneous power and minimized by optimization and control methods.
508

Among the three criterions, the last one is the hardest t o achieve for
anthropomorphosis can not be estimated by equations and principles eas-
ily. Some artificial intelligent methods have been proposed to present the
human walking characters.
These three criterions can also be used t o estimate control methods for
biped mechanism locomotion.
The first two criterions are combined with statistical method in this
paper to evaluate gaits in term of desired ZMP trajectories. Gaits are firstly
converted to ones with the same length of pace, maintaining their individual
speed. Then ZMP trajectories are rated from the view of stability and
energy efficiency point. Using statistical method, desired ZMP trajectory
region with confidence interval will finally be given.
The rest of this paper is organized as follows. Section 2 introduces the
ZMP and energy consumption concept for gait estimation. Based on it, de-
sired ZMP region construction is proposed in Section 3. Related experiment
result is shown in Section 4, followed by the conclusion in Section 5.

2. Formulation of desired ZMP trajectory

2.1. ZMP concept
ZMP is defined as the point on the ground where the total moment gener-
ated due t o gravity and inertia equals zero [12]. It is originally proposed by
Vukobratoivc and Juricic in 1969, and has been widely applied, frequently
cited [13].
Early work related with ZMP criterion minimized the deviation between
the ZMP and the center of shape of the supporting area by treating the po-
sition of the body as free variables [9].Based on the off-line algorithm, a fast
online motion pattern generation technique is adopted in humanoid robot
H7 t o track the desired ZMP when carrying objects [7]. Besides being used
as the optimization objective, ZMP is also applied in constraints design [4].
A convex optimization algorithm for whole body stabilization is introduced
to find the optimal acceleration profile subject t o ZMP constraints [14].
ZMP position can be determined in case of real walking mechanisms
with pressure or force sensors. Dynamical model is required if only partial
counter forces can be measured by pressure or force sensors. Compared
with it, 6-axis force/moment sensor gets practical counter force more di-
rectly and conveniently. It fits for walking on rough terrain or stairs when
both feet are not contacting in the same horizontal plane. Its practical ap-
plication in humanoid robot includes WABIAN-2R [15], Honda [12] and
 509

so forth. Moreover, 6-axis force sensing footwear has been developed for
natural walking analysis in humanoid robot H7 [16].
When sensor is unavailable, ZMP position along x-axis and y-axis can
be calculated by Equations (1) and (2) respectively.

where mi is the mass of link i and g is the gravity acceleration. Position

of link i is described by p = ( x i ,y i , z i ) . And correspondingly, accelerations
of link i in x , y and z direction are represented by xi, yi and Z i . ( I i ) z and
(Ii), are the inertial components. ( & ) , and (&), are the absolute angular
velocity component around x and y axis at the center of gravity of link i.

2.2. Energy consumption calculation

By assuming that power regeneration is unavailable by motor doing negative
work, power can be computer by the product of the motor torque and the
angular velocity as Eei = q q i , where ri and qi are torque and angular
velocity of joint i, i = 1,2, ...,N q . Let T be the time for one gait cycle. We
get the average mechanical power

(3)

to measure the power consumption.

3. Formulation of desired ZMP trajectory using statistical

method
From the calculation equations of ZMP position (Equations (1) and (2)), it
can be found that ZMP position changes with position p and acceleration
p. That is, the desired ZMP trajectory P d z m p satisfies

Pdzmp = f(P,P).
Studies on optimal forward velocities and step length for both nor-
mal adults and biped robot was presented by experimental statistical

following relationship with leg length Lh as L, = (1.0 -

method [17]. For example, desired stride length in human walking has the
1.2) x 15h.l~It
510

is reported that, with given walking speed, there is an optimal step length
that acquires minimal power consumption in both human and robot walk-
ing. Such a walking speed at average pace size is called "natural walking
speed" [ 191.
Similar to that, normal region of pace length and the desired ZMP
trajectory Pdzmp for given robots can be formulated by statistical methods.
Before statistical operation, gait lengthes are converted to standard ones
L, = 1.0 * Lh. Correspondingly, ZMPs are recalculated with individual
velocities.
Let

calculate the average ZMP margin, where pzmpand pb are ZMP position
and stable region boundary.
With Equations (3) and (4), gait i can be estimated by
f i = PeN(I/E.e) + PdN(Ed). (5)
Where N is the normalization operator, be and P d are weighting parameters,
Pe + P d = 1, i = 1 , 2 , ..,N,, N , is the number of gaits. For simplicity,
integration in Equations (3) and (4) is implemented by summarization on
the N , unified samples in one gait cycle.
ZMP trajectories are rated by ZMP margin and energy consumption.
Mean U = [Uz,U,] and covariance V = [Vz,V,] of ZMP probability distri-
bution are given by

4. Experimental Results
Fig.1 is the biped robot Robo-Erectus Senior. After many stages of design
optimizations for its mechanical structure, electronics and control systems,
the final physical parameters of Robo-Erectus Senior are roughly 150cm in
height and 30kg in weight. Each leg consists of a hip, a knee and an ankle
joint, formed by three, one and two motors respectively. These joints are
identical in structure and are composed of a servo motor, a harmonic drive
gear and an incremental encoder. The energy supply, the power circuits
for the joint-drives as well as the micro controller for the low level robot
 511

control are on-board of the robot. Robo-Erectus Senior is equipped with a

monocular USB camera with miniature lenses.

Fig. 1. Robo-Erectus Senior (middle one) and Robo-Erectus Junior (left and right ones)

Ng = 10 experiments have been taken on this platform using following

parameters, Pe = Pd = 0.5, L h = 0.7, N, = 100. Joint torques and ZMP
trajectories are calculated with feedback parameters.
Fig.2 is the Ng ZMP trajectories (see blue * mark) and their weighted
average results. For comparison, the proposed weighting method (see Equa-
tion (5) and red o mark) and fi = (see purple D mark) are applied on
ZMP statistic. It can be found that, statistical ZMP result using average
weights comes to be more compact because that stable gaits with little
energy consumption act as those with much energy consumption in this
statistic method. While the proposed weighting method emphasizes func-
tion of those gaits which are both stable and energy effective.
Let ti be the unified samples in one gait cycle, i = 1 , 2 , ...,N,. Take tl
as an example, Fig.3 is the covariance distribution of V,(t50) and V,(t50).
According to the definition of gait estimation as shown in Equation (5),
the larger the covariance is, the more unstable and energy consuming the
gait is. In words, gaits with smaller covariance tend to be more stable and
512

-0.07 -
*
-0.08 -

-0.09 -

-0.1 -
.--.
-0.11 -
-0.12 - f t%$W
a%*- **
E
-0.13 - tfZ

-0.16'
-0.08 -0.06 -0.04 -0.02 0 0.02 0.04 0.06 0.08 0.1
x(m)

Fig. 2. ZMP trajectories and statistical results

energy efficient,

-00755 -00755 -00754 -00754 -00753 -00752 -00752 -010751 -00751

Y

Fig. 3. Covariances of x and y coordinates at the 50th sample moment

Based on it, Fig.4 shows V(t50)in the region of support foot as -Lb <
x < La and -Wf< y < 0, where L a , Lb and Wfare length of frontal foot,
heel and width of foot.
Thereby, for given V ( t ) ,e.g. V(t50)= 0.006, desired ZMP region can be
found with the help of the statistic result, which is the cone-shaped area
under the partition plate in Fig.4 for the given example.
 513

Fig. 4. Covariances at the 50th sample moment

5 . Conclusion
This paper proposed an experimental construction method of desired ZMP
region based on statistics. Gaits are converted to ones with standard pace
length. Dynamical stability in term of ZMP criterion and energy consump-
tion are estimated on gait samples. Taking the weighted average of gaits as
their means, covariance can be calculated to measure how good the posi-
tion is. From the statistical view of point, region with covariance under the
given value can be treated as the desired area. This result will be used to
generate desired ZMP trajectory for gait optimization and analysis in our
future work.

References
1. S. Kagami, T. Kitagawa, K. Nishiwaki, T. Sugihara, M. Inaba and H. Inoue,
Autonomous Robots 12,71 (2002).
2. M. Vukobratovic and B. Borovac, Int. Journal of Humanoid Robotics 1, 157
(2004).
3. A. Goswami, T h e International Journal of Robotics Research 18,523 (1999).
4. M. Morisawa, S. Kajita, K. Kaneko, K. Harada, F. K. andK. h j i w a r a and
H. Hirukawa, Pattern generation of biped walking constrained on parametric
surface, in Proceedings of the I E E E International Conference o n Robotics and
Automation, 2005.
5. N. F. Troje, Journal of Vision 2,371 (2002).
6. G. Mather and L. Murdoch, Gender discrimination in biological motion dis-
514

plays based on dynamic cues, in Proc. of the Royal Society of London, B:

Biological Sciencesl994.
7. K. Nishiwaki, S. Kagami, Y. Kuniyoshi, M. Inaba and H. Inoue, Online gen-
eration of humanoid walking motion based on a fast generation method of
motion pattern that follows desired zmp, in IEEE/RSJ International Con-
ference on Intelligent Robots and System, 2002.
8. Y. Zhang, Q. Wang, W. Qiang and P. Fu, A new method of desired gait
synthesis in biped robot, in Proc. of the 3rd World Congress o n Intelligent
Control and Automation, 2000.
9. C.-L. Shih, Y. Zhu and W. A. Gruver, Optimization of the biped robot tra-
jectory, in Proc. of IEEE Int. Conf. on Systems, Man, and Cybernetics, 1991.
10. Q. Huang, K. Yokoi, S. Kajita, K. Kaneko, H. Arai, N. Koyachi and K. Tanie,
IEEE Trans. on Rob. and Aut. 17,280 (2001).
11. H. Elftman, Journal of Bone and Joint Surgery 48A, 363 (1966).
12. K. Hirai, Current and future perspective of Honda humanoid robot, in Proc.
of IEEE Int. Conf. on Intelligent Robots and Systems, 1997.
13. M. Vukobratovic and D. Juricic, IEEE Trans. o n Bio-Medical Engineering
16, l(1969).
14. J. Park and F. C. Park, Advances in Robot Kinematics (Springer Nether-
lands, 2006), ch. A convex optimization algorithm for stabilizing whole-body
motions of humanoid robots, pp. 157-166.
15. http://www.takanishi.mech.waseda.ac.jp/research/wabian/index.htm.
16. Y. Takahashi, S. Kagami, Y. Ehara, M. Mochimaru, M. T . M and H. M.
H., Six-axis force sensing footwear for natural walking analysis, in IEEE
International Conference on Systems, Man and Cybernetics, 2004.
17. M. Hardt, 0. von Stryk, D. Wollherr and M. Buss, Development and con-
trol of autonomous, biped locomotion using efficient modeling, simulation,
and optimization techniques, in Proc. of IEEE Int. Conf. on Robotics and
Automation, (Taiwan, 2003).
18. E. A. B. John, P. D’ANTONIO, J. PARD0 and V. L. David, Journal of
motor behavior 34,309 (2002).
19. L. Roussel, C. Canudas-De-Wit and A. Goswami, Generation of energy o p
timal complete gait cycles for biped robots, in Proc. of IEEE Int. Conf. on
Robotics and Automation, 1998.
 LOCOMOTION CONTROL SCHEME FOR FAST WALKING
HUMANOID SOCCER ROBOT

WEERAWT SAWASDEE
Institute of Field Robotics (FIBO), King Mongkut 's University of Technology Thonburi,
Bangkok 10140, Thailand

PASAN KULVANIT
Department of Science Service, Ministry of Science and Technology, Bangkok 10400,
Thailand

THAVIDA MANEEWARN
Institute of Field Robotics (FIBO), King Mongkut 's University of Technology Thonburi.
Bangkok 10140, nailand

In the game of soccer, the ball approaching speed of the humanoid robot is vital. Fast
moving robot gains advantage over the other players. For a limited computing resource
and actuator specifications our kid size humanoid robot can walk fast with robustness to
external disturbance in the form of stepping foot placement impact. The inverted
pendulum model is used for walking pattern generator. The step size and sway distance
are adapted online using feedback information from gyro sensors. The impact reduction
scheme based on foot force feedback information is implemented to damp out excessive
impact force at heel strike. The locomotion scheme is implemented successfully with our
humanoid soccer robot competing in the Robocup 2007 Humanoid Soccer League.

1. Introduction
Most biped robot systems nowadays has very little problem stabilizing the
walking motion on an ordinary flat surface [l, 2, 3, 41. The next step may be
finding the way to increase the walking speed or achieving the running motion.
Fast walking speed or running motion helps the robot reaching the destination
faster and renders the humanoid robot more useful than a normally slow pace
robot. In humanoid soccer match, it is obvious that the robot with fast ball-
approaching speed can reach the ball faster than its opponent and be able to
shoot or snatch the ball giving the team more chance to score the goal or
defending its own goal.
The problem of walking fast is the effect of the inertia force to the dynamic
stability condition is more pronounced. The foot placement scheme to match the

515
516

ground reaction force and the planned Zero Moment Point (ZMP) must be
realized and the compensation must be computed in real-time. The formal way
of calculating Z M P [8] may not satisfy the real-time requirement here due to
computing hardware and time limitation.
The other problem stemming from the fast walking robot is the impact
created by the heel strike event. This impact force can cause the humanoid robot
to deviate from planned trajectory or even fall down completely. The
experiment results from [2] and the data from real human walking from [9]
show that the impact force at heel strike increase proportionally with the
walking speed. In normal human walking with the speed of 2 to 4 Kilometer per
hour, the impact force is approximately 1.2 to 1.4 times the body weight. If the
walking speed increases to 8 Kilometer per hour, the impact force is
approximately 1.8 times the body weight. Human body is structurally build to
deal with such impact force. The soft tissue around the foot and the tendon
structure of the foot help damping out such impact disturbances. For the bipedal
walking machine with mechanical body equipped with multiple DC servo
motors and no passive energy storage devices, there is no help from compliancy.
The introduction of compliance into the rigid system is needed. Park [7]
suggests the approach of using the software to help damping out the impact
force. [2, 61 use simple compliance in the form of flexible pad andor springy
foot mechanism attaching to each foot of the humanoid robot in order to
partially damped out the vibration created by the impact.
This work shows the small size and compact humanoid robot system that
can walk stably fast on the nominal flat surface with ability to treat the impact
force that may de-stabilize the walking locomotion. There are three main
controllers that help stabilizing the walking pattern. First is the regulation of the
body uprightness using the feedback information from gyro sensor. Second is
the adaptation of the step size based on the concept of the ZMP. The observed
value of the angular velocity is used here to predict the direction of the robot
CM's acceleration and adjust the foot placement position. Third is the controller
that deal with the impact force via the vertical foot force feedback. The
compliance at the knee is adapted with software using foot force information.
All three controllers work collectively to adapt the pre-planned walking pattern
in real time. Note that all schemes are executed with a low-power consumption
and low cost microcontroller. Comparing to the more complicated humanoid
robot systems such as [ 1, 2, 31 the resulting walking performance is quite good.
The robot is used in the Robocup 2007 Humanoid Soccer League Robot
competition [5] with satisfying results.
The structure of this paper is as follows. Section 2 presents the biped robot
system used to obtain the experiment results. Section 3 discusses the locomotion
control and the impact reduction scheme. Section 4 shows the results of
implementation on the biped robot. We then conclude this paper with section 5.
 517

2. Humanoid robot system

The actual biped robot design and build at the Institute of Field Robotics (FIl30)
is used in the experiment process. The general specifications are displayed in
Table I. The robot is originally built to compete in the Robocup Humanoid
league competition.

Table 1 Biped robot system specifications

Name: Jeed Specifications

Height 45 CM
Weight 2.7 KG (total)
Walking Speed 18 d m i n (max)
Number of 22
DOF
Actuator DC servomotor
Structural Aluminum Alloy
Material

There are 2-axis accelerometer [+/-2g], 2 rate gyros [+/-I00 deglsec], 8 force-
sensing resistors, and a camera. The accelerometer tells the robot if there is any
longitudinal and/or transversal tilt. The two rate gyros, installing at the lower
trunk,measure angular velocity at longitudinal and transversal axis. The angular
velocity information will be used to adapt the attitude of the body during
walking and calculating the instantaneous acceleration of the CM. The force
sensing resistors are used to sense the vertical ground reaction force. Force
information will be used to adapt the knee joint impact absorption characteristic.
The camera is used to track the ball and other objects of interest, which is
crucial for navigation decision-making software.

8233

Figure 1 . System overview of the biped robot

Figure 1 shows the overview of the biped robot system. The main
computer for the biped robot is ARM-7 [60MHz] RISC microprocessor, Its
major function is to make decision based on the information obtained from the
518

camera system and sensors. The main computer, then, issues commands to the
motor locomotion controller (the other ARM-7 processor) to generate walking
motion. The inverse kinematics of the robot legs is stored in ARM-7 motor
controller. The kinematics is calculated in real time.

3. Locomotion control and impact reduction scheme

The small size and compact humanoid robot system can walk stably fast with
the ability to treat the impact force that may de-stabilize the walking locomotion.
There are three sets of the control scheme which are designed to help stabilizing
the walking pattern at before, during and after the foot placement as shown in
figure 2.

-g
Paametas
s-g

m g TYPC
Selection
v
-g
Pattun
Gmaauon
c
lnvase
KmaMtlcr
Impact
Reducllon
Controller
PorltlDn PD controller

hguk vdocuy %ct

to Robot Ddecllon
A

Figure 2. Locomotion control scheme

3.1. Before foot placement

3.I . 1 Controlling balance before foot placement

When the robot loses its balance, the angular velocity can be sensed from the
gyro sensors. The measured angular velocity is fed back to the system so that
the robot can lean its body to compensate for the angular rotation which can
help balancing the robot before the foot placement. The angular position of four
motors attached to the robot's ankle are adjusted directly from the sensed
 519

angular velocity. When the inverse kinematics is used to calculate joint positions
from the predefined gait which is used as the reference trajectory, the PD
controller is used to adjust the position command which is the input for these
motors at the ankle as shown in Eq.( 1)

CMD = CMD f((Kp(Erq))+ ((Erq-,- Erq)Kd)) (1)

CMD is the command angular position

Kp9Kd is the PD gain
Erq is the angular velocity error at this time step
Er5-1 is the angular velocity error at the previous time step

3.1.2. Controlling the foot placement

When the robot walks, inertial force and gravitational force affect the walking
acceleration. These forces can be referred to as the total inertial force. When the
foot touches the ground, the robot receives the reaction force from the ground.
The point of intersection between the ground reaction force vector and the total
inertial force vector is called the zero moment point because it has zero moment.
The position on the ground that the reaction force passes through is called the
ground reaction point. When the walking pattern is generated, the target total
inertial force can be calculated from the robot model. When the robot can
perfectly balance during the walk, the target total inertial force and the actual
ground reaction force is projected to the ground at the same position. When the
robot walks on an uneven ground , these two positions will be separated which
results in lost of balance and causing the robot to fall. The force that causes the
fall can be shown in term of the mismatch Z M P and the actual ground reaction
force as shown in figure 3.

Target ZMP=Center ofGround Reaction Target ZMP Center of GroundReaction

figure 3. the diagram shows the robot losing balance

When the robot is starting to fall forward, the velocity and acceleration increase
in the forward direction. The ZMP control system can help the robot's body
position to move toward the direction of acceleration by adjust the foot
520

placement position. The ZMP control system is designed to adjust the walking
step size based on the relationship with the predicted direction of the body
acceleration. The direction of the body acceleration is predicted from the sensed
angular velocity of the robot's CM. The value and direction of the sensed
angular velocity of the robot's CM (in two axes) are scaled and then added to the
planned foot placement position in x and y axis. The adjusted foot placement
compared to the planned foot placement are shown in figure 4.

3.2. During footplacement

The impact reduction control is used to reduce the high impact force occurred
during foot placement. The force that is acting to the foot when the foot touches
the ground can be measured by force sensors that attached to the four corners of
the robot's foot. The impact force can be identified from the average force from
all four sensors. When the value of force and the average change of force is
higher than the specified threshold in the positive direction, the impact state is
established. In the impact state, this average change of force is used for the knee
bending adjustment. The knee bending adjustment depends on the amount of
impact force. The difference between the average change of force and the
specified threshold is scaled (in the P-control fashion) and this amount is then
taken out from the height of the robot which is the input for the knee bending
inverse kinematics equation. The command input for the motors attached to the
leg are then adjusted accordingly.

3.3. After footplacement

The control algorithm for after impact is similar the PD controller that control
the angular velocity before the impact. When the angular velocity after impact is
controlled by the PD controller, the controller can successfully reduce the
velocity to be within the acceptable level and stabilizes the walking cycle. When
the PD controller is not used, the robot tends to falling forward after the impact.

4. Results
We implement the locomotion control scheme to the real humanoid robot. The
robot walks at the nominal speed approximately 0.2 m / s . We detect the
acceleration at the foot via vertical accelerometer as shown in figure 5. The
angular velocity in the side sway direction of the CM is measured from the side
sway gyro sensor. We can see that when the impact reduction scheme is off,
 521

there are many sub-peaks occur in the vertical acceleration profile at the foot.
When the impact reduction scheme is on, the sub-peaks are partially eliminated
and only left with the major peaks, which represent the initial impact at heel
strike. Note that the unit of the time axis is second and the unit of the vertical
axis is volt since we map the force quantity to voltage.
V V

3.0

2.0

1.0

t o t
0.5 1.o 1.5 0.5 1.0 I5

Figure 5. Vertical acceleration at the foot during stepping motion. Left figure is the acceleration for
the case without the impact reduction scheme. Right figure is the acceleration profile for the case
with the impact reduction scheme.

We compare the result of the side sway behavior of the humanoid robot between
the case when gyro feedback control is turned on (Figure 6. Right Figure) and
the case when gyro feedback control is turned off (Figure 6. Left Figure). The
graph of the gyro reading of the side sway behavior is periodic. This means that
the controller is acting to stabilize the swaying motion. The left figure shows no
constructive pattern, which means that the swaying motion is not rhythmic and
the robot is now prone to falling down. Note that the unit of the time axis is
second and the unit of the vertical axis is volt since we map the angular velocity
to voltage.

V V

5.0 5.0

4.0 k.0

3 .O 3.0

a .o 3.0

1.o 1.0

". ~~ - -~~~ -- "

1.0 2.0 3.0 k.0 1.0 1.0 3.0 4.0

Figure 6 . Angular velocity data read from side sway gyro sensor. The right figure is the case when
the gyro feedback control is on. The left figure is the case when the gyro feedback control is off.

5. Conclusions
Fast and simple controllers are introduced to the humanoid robot locomotion
control scheme. The controllers can be divided into three main parts. First is the
522

preview control to generate the foot placements for the fast walking motion.
Second is the gyro feedback control to help regulating the trunk of the robot to
be upright during walking. Last controller is the impact reduction control
scheme to help damping out the excessive impact force at the heel-strike event.
All three controllers work together to stabilize the bipedal walking as a whole
and help increasing the robustness of the walk. The result is the fast and stable
biped locomotion that is suitable for humanoid soccer robots.

Acknowledgements
We would like to thank Asian Honda Motor Co., Ltd., Chevron (Thailand) Ltd.,
Advance Info Service PLC., and The Engineering Institute of Thailand for their
generous support Team KMUTT in the Robocup Soccer Robot competition.

References
1. Hirai, K., Hirose, M., Haikawa, Y., and Takenaka, T., “The development of
Honda Humanoid Robot”, Proceedings of the 1998 IEEE International
Conference on Robotics &Automation, V01.2, pp. 1321-1326 (1998).
2. Honda Motor Co., Ltd., Public Relations Division, “ASIMO, The Honda
Humanoid Robot”, Technical Information (2002).
3. Kagami, S. et al., “Online 3D Vision, Motion Planning and Bipedal
Locomotion Control Coupling System of Humanoid Robot: H7”,
Proceedings of the 2002 IEEE/RSJ International Conference on Intelligent
Robots and Systems, Vol. 3, pp.2557-2562 (2002).
4. Kajita, S. et al., “The 3D Linear Inverted Pendulum Mode: A simple
modeling for a biped walking pattern generation”, Proceedings of the 2001
IEEE/RSJ International Conference on Intelligent Robots and Systems, Vol.
1, pp. 239-246 (2001).
5. Kulvanit, P., et al, “Team description paper: Team KMUTT”, Robocup
Humanoid League 2007 (2007).
6 . Nagasaki, T., et al, “A Running Experiment of Humanoid Biped”,
Proceedings of 2004 IEEERSJ International conference on intelligent
robots andsystems, Vol. 1, pp. 136-141 (2004).
7. Park, J.H., “Impedance Control for Biped Robot Locomotion”, IEEE
Transactions on Robotics and Automation, Vol. 17, No.6, pp. 870-882
(2001).
8. VukobratoviC, M, Boravac, B., Surla, D., and Stokic, D., Biped Locomotion
ScientiJic Fundamentals of Robotics 7, Springer-Verlag (1989).
9. Zatsiorsky, V.M., Kinetics of Human Motion, Human Kinetics (2002).
 OPTIMUM PERFORMANCE OF THE FAST WALKING
HUMANOID SOCCER ROBOT: EXPERIMENTAL STUDY

PASAN KULVANIT
Department of Science Service, Ministry of Science and Technology, Bangkok, 10400,
Thailand

BANTOON SRISUWAN
Institute of Field Robotics, King Mongkut 's University of Technology Thonburi,
Bangkok, 10140, Thailand

DJITT LAOWATTANA
Institute of Field Robotics, King Mongkut 's University of Technology Thonburi,
Bangkok, 10140, Thailand

The optimum performance of the humanoid robot is needed in the game of robot soccer.
In order to gain advantage over the opponent, the robot must output the fastest possible
walking speed at the minimal expense of on-board power source. We implement the gyro
feedback control to the inverted pendulum model to regulate the attitude of the trunk
during walking. Through experimental study with the real humanoid robot, we have
found the optimum step size the humanoid robot should take in order to walk fast at
minimum power consumption. Our robot is able to achieve the optimum speed of 0.18
m/S.

1. Introduction
A humanoid robot is composed of various sub-systems working together to
achieve a controlled bipedal walking locomotion. Such a complex system is
difficult to design exactly to the required specifications. Only estimated
performance can be predicted. The optimal performance can only be obtained
via experiment and testing. The needed performance for humanoid soccer robot
is maximum walking speed and minimum power consumption. Walking pattern
with lengthy double support phase such as in [3, 101 are not sufficient for soccer
playing robot. [9] realize a natural walk using angular momentum information.
[2] state that although the system of inverted pendulum is unstable in the sense
of Lyapunov, we can still design a stable trajectory for the gait by switching the
support leg at the mirror location of the angular position of the starting point of
the center of mass (CM)'s trajectory. The condition for leg switching follows

523
524

the locomotion continuity condition where the terminal CM’s velocity of

previous cycle equals the CM’s initial velocity of the present cycle. [4]
comment on the phase plane trajectory of the constant hip height walk that a
kind of energy is conserved and called it orbital energy. The behavior of the
trajectory is much depended on the orbital energy. In addition, the constant
horizontal hip velocity minimizes “jerk” in the trunk area. Minimization of jerk
and the hip ripple increases walking reliability due to decreasing in a risk of
encountering unexpected external disturbances.
In this paper, we use a constant hip height, gravity compensated inverted
pendulum model to construct a gait pattern. The energy and angular momentum
of such gait is conserved. The gait also has self-stabilized trajectory in nature.
The gyro feedback control is used to adapt the gait pattern in real time in case
when the robot encounters any disturbances. The experiment is used to find the
optimal operating point, i.e., natural dynamic, for the robot.
The structure of this paper is as follows. Section 2 introduces the humanoid
robot system used for this work. Section 3 explains the inverted pendulum
model used to implement fast dynamic gait and the adaptation of gait pattern via
gyro stabilization control. Section 4 presents the findings from experiment.
Section 5 concludes the paper.

2. Humanoid robot system

The actual humanoid robot, which is designed and built at the Institute of Field
Robotics (FIBO) is used for the experiment study. The robot competes in the
Robocup Soccer Humanoid League 2007. Table 1 shows the overall
specifications of the robot.

There are 2-axis accelerometer I+/-2g], 2 rate gyros [+/-lo0 deghec], 8

force-sensing resistors, and a camera. The accelerometer tells the robot if there
 525

is any sagittal and/or transversal tilt. The two rate gyros, installing at the lower
trunk, measure angular velocity at sagittal and transversal axis. The force
sensing resistors are used to sense the vertical ground reaction force. The
camera is used to track the ball and other objects of interest in the soccer field.

Figure 1, System overview of the biped robot

As shown in Figure 1, the main computer for the biped robot is ARM-7
[~OMHZ] FZISC microprocessor. Its major function is to make decision based on
the information obtained from the camera system and sensors. The main
computer, then, issues commands to the motor locomotion controller to generate
walking motion. The inverse kinematics of the robot legs is stored in ARM-7
motor controller. The kinematics is calculated in real time.

3. Fast dynamic gait

In our fast dynamic gait, we minimize time spent on an exchange of support
phase and keep the biped robot stable by relying on the high gain control at the
single support phase. By switching the leg fast enough, the change of support
phase is squeezed into a very short time in the gait cycle such that it produces
infinitesimal effect on locomotion’s stabilization process. The inverted
pendulum model is the model of choice. It can be readily described by the four
basic walking parameters - Step length/ Side sway distance/ Trunk bend angle/
Step time interval.

3.1. Two dimensional Inverted pendulum model

We, first, pay attention to the single support phase of the bipedal walking, where
the walking cycle occupy the largest percentage of the total walhng. As shown
in figure 2(a), we represent the bipedal walking during the single support phase
526

as a variable length inverted pendulum. Assume that all biped robot’s mass is
concentrated at the top most part of the pendulum. Let’s assume further that
there is external torque acting at the ankle. In real robot, this is the torque that is
produced from the power ankle joint.

(a> (b)
Figure 2: a) Variable length inverted pendulum model representing the single support phase of
bipedal walking b) A free body diagram corresponded to the variable length inverted pendulum
model
Figure 2(b) shows the free body diagram of all forces and torque that act on the
CM of the inverted pendulum. The equations of motion are:
Rotational dynamic

Linear dynamic

? - 0‘ 2 r+gcosB=- f
M
Since the vertical component of the force acting on the CM are fcose and
Mg; therefore in order to cancel out gravity effect, f must equal Mglcose. We
can, then, achieve a constant hip height trajectory. Figure 3 shows the phase
plane plot of the inverted pendulum motion with constant hip height under the
influence of several sets of initial conditions. Note that the ankle is not active.
We use a set of initial condition that consists of initial angular position (8)and
angular velocity (8).There are nine sets of initial conditions. Some of them
manage to travel from negative angular position to the positive region. Some of
them cannot make it to the positive side. The phase plane plot shows that the
system of inverted pendulum is unstable since the trajectory of the system
neither goes to the equilibrium point nor constitutes a limit cycle. However, the
stable trajectory that satisfies the locomotion continuity condition can be
achieved. As shown in Figure 3, the 4‘htrajectory from the top has equal value
of at the beginning and at the end of the trajectory. We will use inverted
 527

pendulum model as a basis to generate a walking pattern for the humanoid

robot.

1 08 06 -04 02 0 02 04 06 08 1
Theta

Figure 3 Phase plane plot for the inverted pendulum model mass in motion at constant hip height
(i.e., gravity compensated motion)

3.2. Gyrofeedback control scheme

The walking pattern based on gravity compensated inverted pendulum model
alone is not sufficient to stabilize the walking robot. We need adaptation scheme
to compensate the error that happens due to external disturbance. We use two
gyro sensors to feedback the attitude information in 2 axes- sagittal and
transversal plane. This information tells the walking controller the states of
angular velocity at each sampling moment. We sample the gyro sensor data
every 10 ms. The control loop for bipedal locomotion used with the real system
is shown in Figure 4. Assuming that this controller is controlling the attitude of
the biped robot in the sagittal plane. In t h s case the angular velocity of the
ankle’s joint in the sagittal plane is regulated based on the feedback information
from the sagittal plane’s gyro sensor. We use the ankle torque to control the
whole body attitude of the biped during walking. The control equation is as
follows.
Wankle-cmd = K p (wgyra - Wkinematics -k Wkinemalics (3)
Where
Wankle - cmd
: Angular velocity command for the ankle joint’s motors
KP : Proportional gain
- : Angular velocity reading from gyro sensor
0
,

Wkinemarics
: Joint angular velocity calculated from the inverse kinematics
528

Figure 4: Bipedal motion control loop used to stabilize the fast gait

The walk control scheme of [lo] involved the following torque control
equation:
- Lmhr 1+ T4
Tre/ = KL Ge/ (4)
Eq. (4) is the feedback control scheme used to regulate the angular momentum
of the robot ( Lrobot) to some reference value ( Lrer)by way of adapting the

ankle torque (c). If we let K , = K , / A t and we know that T = Io, then

eq. (3) is equivalent to eq. (4) and the angular momentum of the robot with the
gyro feedback control scheme is conserved.

4. Experiment results
We operate the robot at three different speeds- 0.12, 0.16, and 0.18 d s . For
each walking speed we vary the step size and tune other walking parameters to
stabilize the walk. The total power consumption is recorded via load current
sensor. The walking speed is the average walking speed.
Figure 5 shows three sets of data at three different CM’s speeds. Note the
parabolic nature of the plot. The trend of each data set is that, at the same
walking speed, the energy consumption takes parabolic characteristic such that
at the shortest step length and longest step length, the energy consumption is at
maximum. There is an optimal step length that the biped robot can use to
acquire minimal power consumption. It should be noted that the biped that walk
with higher speed tends to consume more power than the biped robot walking at
slower speed. Some of the data points that distributed out of the parabolic range
are the results of some discrepancy during the practice of the experiment, and
assume to be ignored. It can be concluded that for the biped robot walkmg with
the same speed, there is an optimal step length that leads to optimal power
consumption. For our robot used in the experiment, such step length is
approximately 0.046 m. This is also true for human locomotion. Such a walking
speed at average human step size is called “natural walking speed” [8] or “self-
 529

selected speed”. This natural wallung speed is the walking speed that
minimizes the human’s power consumption.

Step length (mrn) 3

Figure 5. For the biped walk with the same CM’s speed, there is an optimal step length that
minimizes the power consumption used during the walk. Grey dots represent the walk at 0.18 d s .
Squares represent the walk at 0.16 d s . Solid dots represent the walk at 0.12 m/s

We select 0.18 d s as our operating speed. Comparison via Froude number,

Fr = v2/(g*Lleg),where v is walking speed and Llegis leg length [5], shows that
our robot (Fr = 0.018) performs equally well when compared with WABIAN-
2R [7] (Fr = 0.021) and performs better than HRP-2LR [6] (Fr = 0.0043) or
Hajime Robot [ l ] (Fr = 0.0136). It should be noted that the maximum speed of
our robot is 0.304 d s with higher power consumption than the one at the
optimum walking speed.

5. Conclusions

An experimental study of the humanoid robot to achieve an optimal

performance is the focus of t h s work. The interested objectives are walking
speed and power consumption. The basic gait pattern for our robot is the gravity
compensated inverted pendulum model. The implementation of the gyro
feedback controller is for added locomotion stability robustness. The result
obtained is the step length of 0.046 m that always yield minimum power
consumption at any walking speed. This result is specific to the particular robot.
It is crucial to find this “natural parameter” in order to operate the biped robot at
the optimum level.
530

Acknowledgments
This research is partially supported by the Thailand Research Fund through the
Royal Golden Jubilee Ph.D. Program (Grant No. PHDA .M.KT.44/C.1). We
would like to thank Asian Honda Motor Co., Ltd., Chevron (Thailand) Ltd.,
Advance Info Service PLC., and The Engineering Institute of Thailand for their
generous support for Team KMUTT in the Robocup Soccer Robot competition.

References

1. Friedmann, M., et al., “Team description paper: Darmstadt Dribblers &

Hajime Team (Kidsize) and Darmstadt Dribblers (Teensize)”, Robocup
Humanoid League 2006 (2006).
2. Gubina, F., Hemami, H., and McGhee, R.B., “On the Dynamic Stability of
Biped Locomotion”, IEEE Transaction on Biomedical Engineering, Vol.
BME-21, No.2, pp. 102-108 (1974).
3. Ito, S., Asano, H., and Kawasaki, H., 2004, “A Weight Shift by Control of
Center of Pressure of Ground Reaction Forces in Biped Double Support
Phase”, Journal of the Robotics Society of Japan, Vol. 22, No.4, ppl13-
120.
4. Kajita, S., Tani, K., and Kobayashi, A., “Dynamic Walk Control of a Biped
Robot along the Potential Energy Conserving Orbit”, International
Workshop on Intelligent Robots and Systems, Vol. 2, pp789-794 (1990).
5. McNeil, A.R., “Gaits of Mammals and Turtles”, Journal of the Robotics
SocietyofJapan, Vol.11 No.2,pp 314-319 (1993).
6. Nagasah, T., et al., “A Running Experiment of Humanoid Biped”,
Proceedings of 2004 IEEEIRSJ International conference on intelligent
robots andsystems, Vol. 1, pp. 136-141 (2004).
7. Ogura, Y., et al., “Evaluation of Various Walkmg Patterns of Biped
Humanoid Robot,” In proceedings of the 2005 IEEE International
conference on Robotics and Automation, pp. 603-608 (2005).
8. Roussel, L., Canudas-de-wit, C., and Goswami, A., “Generation of Energy
Optimal Complete Gait Cycles for Biped Robots”, Proceedings of the 2998
IEEE International Conference on Robotics and Automation, Vol. 3, pp.
2036-2041 (1998).
9. Sano, A. and Furusho, J., “Realization of Natural Dynamic Walking Using
the Angular Momentum Information”, Proceedings of 1990 IEEE
International Conference on Robotics and Automation, Vol. 3., pp. 1476-
1481 (1990).
10. Takanish, A., et al., I., 1989, “Dynamic Biped Walking Stabilized with
Optimal Trunk and Waist Motion”, IEEE/RSJ International Workshop on
Intelligent Robots and Systems, pp. 187-192.
Supporting Technologies
This page intentionally left blank
 A BIOLOGICALLY INSPIRED ARCHITECTURE FOR
CONTROL OF GRASPING MOVEMENTS OF AN
ANTHROPOMORPHIC GRIPPER

s. VARONA MOYA~,J. MOLINA VILAPLANA~,A. LINARES BARRANCO~,J. FELIU

BATTLEa, J. L6PEZ CORONADOa
‘Department of Systems Engineering and Automatics, Polytechnic University of
Cartagena, Campus Muralla del Mar, C/Dr. Fleming S/N 30202, Cartagena, Murcia,
Spain

bDepartment of Computer Technology and Architecture, University of Sevilla, Avda.

Reina Mercedes S/N. 41012 Sevilla. Spain

In this paper we introduce a research plan for the development of an architecture for
control of grasping movements in a manipulation system. It consists in an
anthropomorphic robotic gripper with tactile and propioceptive sensing capability plus a
stereoscopic visual guidance system, electronic control boards designed for implementing
a powerful and neurologically inspired inter-chip communication technique called AER
and a HYPBF neural networks model that will emulate the performance of the most
important brain areas involved in motor behaviour control for efficient object
manipulation.

1. Introduction
How does the brain control our hand when we use it to grasp something? As
with any other complex system, the initial approach to answering this question
starts with a black box model where inputs and outputs are mapped in a
convenient way. We know several thmgs about this box: it is located in the
central nervous system, inputs to the box must be visual, tactile, propioceptive
and task-related flows of information, the output is a prehension behaviour
measured in terms of static gestures and dynamic forces during task execution
and the end-effector gripper can be a real human hand, a prosthesis or a robotic
mechanism. Also, knowledge obtained from previous grasping operations must
be stored and recovered inside that black box.
Following MacKenzie [ 11, several disciplines have tried to model this
system, such as control engineering, robotics, mathematics and computer
science. Regarding the engineering approach, human motor control has been
identified as a slow adapting, low resolution, sensor-driven non-linear system.
However, it has not been possible, based on these assumptions, to develop

533
534

systems capable of matching the skills that humans and other animals show
when manipulating things and, above all, speaking of cognitive functions (Arbib
[21)).
It is clear that, at present, there is not a simple solution to this problem.
Even now that we can dispose of high computational power, it is still necessary
to solve certain questions such as efficient coordination of hand joints regarding
synergic movements, development of bio-inspired sensors, design of an
appropriate control of end-effector forces in order to achieve an efficient and
stable grasp in a highly anticipatory manner, also capable of reacting to
unexpected events like object slipping, ways of extracting information about the
object and the task and, finally, the implementation details in order to make
simulations and obtain results.
With the aim of structuring and guiding research work, Marr [3]
suggested three levels in the analysis of the performance of any system that
process a big amount of information in a complex way. These are the task level,
the representation or algorithmic level and the implementation level.
Computational models implement the ideas present in conceptual models, letting
us test them and obtaining useful data that can suggest modifications or even
new conceptual models. Expert systems of artificial intelligence, non-linear
equations and, more recently, neural networks can be found in this level.

Conceptual Models

validate
Computational Models 4 b Experimental Models
suggest

Figure 1. Marr’s [3] triangle strategy.

Nowadays, neurorobotics has emerged as a fresh research field in

which the method of cognitive neuroscience can be applied to robotics in order
to present a bio-inspired solution to the grasping problem. This is one of the
research lines of the NEUROCOR (NEUrotechnology, Control and Robotics)
group of the Polytechnic University of Cartagena (UPCT). Our aim is to
develop biologically inspired control systems that achieve the closest human-
like performance of robots in a wide area, applying the functional principles of
the highly efficient central nervous system, gaining robustness and flexibility.
 535

In cognitive neuroscience, clinical data and brain imaging are used to

form a high level conceptual model of the involvement of various brain regions
in the studied behaviour -grasping motor behaviour, in our case-. As Arbib [2]
details in his brain theory, we may model the brain either functionally, analyzing
some behaviour in terms of interacting schemas, or structurally, through the
interaction of anatomically defined units, such as brain regions, or substructures
of these regions, such as layers or columns, seeking an explanation in terms of
neural networks, elements of Marr’s computational level, since the neuron may
be considered the basic unit of function as well as of structure. However,
schemas may be implemented not only in neural networks, but also in
conventional computer programs or special-purpose devices. All these
approaches rest on effective design of VLSI chips or other computing materials.
Using neural networks in the computational level has the advantage of allowing
the researcher to take direct inspiration in the performance of the nervous
system and exploit, at the same time, the benefits of working with simple yet
powerful computational units that offer a biologically plausible solution.

A 3

Figure 2. Arbib’s [2] views of level of analysis of brain and behaviour (A). , and a distributed
technological system (B), highlighting the role of schemas as an intermediate level of functional
analysis in each case

2. The cortical motor system

Behavorial, cell recording and anatomical data from both human and monkey
have been used to anatomically and functionally identify cortical areas that
control motor functions. As a brief overview, two of the most important areas
are the rostra1 inferior premotor cortex (referred to as F5) and the anterior intra-
parietal cortex (AIP). It is suggested that F5 contains a storage (“vocabulary”)
of motor actions (Rizzolatti and Luppino [4]), whose words are represented by a
536

population of neurons that codify an action at different degrees of abstraction.

It is thought that when an appropriate stimulus is shown, F5 neurons
automatically code a potential action, whch is the representation of that action.
A sector of F5, FSab, receives a major parietal input from area AIP (Luppino et
al. [5]). The study of the responses of area AIP neurons indicates that they code
three-dimensional objects in visual terms. Taken together, the circuit formed by
AIP and F5 plays a role in visuomotor transformations for object grasping and
manipulation (Rizzolatti and Luppino [4]). Various attempts have been made to
explain how AIP and F5 neurons perform that transformation. Common to these
proposals is the idea that some neurons of AIP code the object’s intrinsic
properties and, then, either directly or via other AIP elements, send this
information to specific sets of F5 neurons. F5 transforms the received
information into patterns of hand movements that are appropriate to the size and
shape of the objects to be grasped. Finally, the F5 grasping pattern recruits
specific F1 neurons to execute the grasp (Rizzolatti and Luppino [6]).

3. Computational models for grasping behaviour

Several computational models have been developed following this trend of
research. Not all of them have gone down to the level of modelling those areas,
but they all rely on exploiting the learning capabilities of neural networks as a
way to emulate the distributed information process that is presumed to take
place in them and they have all been inspired at some extent by physiological
evidence. Uno’s model, for instance, tries to solve the grasping problem by
modelling a multilayered neural network that learns mappings between objects
and grasping gestures, storing those bindings as an abstract representation in an
internal layer (Uno [7]). Other models do work with the mentioned brain areas,
like those of Sakata et al. [S], Moussa [9] and above all Fagg and Arbib [lo]
and Rizzolatti and Luppino [ 111 who proposed an alternative model to Fagg and
Arbib’s model.
In addition to these proposals, NEUROCOR is developing a model of
radial basis functions (RBF) based neural networks that performs the dynamic
visuomotor transformation made in F5 and AIP, following the interaction model
presented by Murata et al. [13], as seen in figure 3, based on the idea that
parietal areas of the cortex can correlate different types of sensorial information
or different neural coding in a single space of representation through a process
that Burnod and his partners (Baraduc et al. [ 14,151) called matching process.
From this point of view, the correlation or matchtng between two different
sensorial types or different codings is made in a matching unit. We propose that
AIP can be understood as one of these matching units. Indeed, the closed loop
of the circuit below is a mechanism that allows a continuous comparison
 537

between the internal representation regarding the grasp target (the object) and
another representation regarding the present motor action, generating at every
moment the necessary motor commands until the task is completed.
F1 F5 AIP GIP

al input

Figure 3. Model of interactions between different cortical areas by Murata et al. [13] for control of
the prehension movement.

4. A new control architecture

Based on this neuro-inspired control policy, NEUROCOR is developing and
implementing neural networks to control a ma~pulationsystem composed of an
anthropomorphic robotic gripper and a stereoscopic visual guidance system
(Coronado et al. [16], Taddeucci [17]).
The goal in our approach to efficient object manipulation is to
implement a multisensorial neuro-inspired control architecture. It is new in the
sense that visual, tactile and propioceptive information will be integrated and
handled together; also, because it will be highly modular and also completely
platform independent. Its control core will be a group of interconnected HYPBF
(hyperplane basic hctions) neural networks that will be ~ c t i o n a l l y
equivalent to certain brain areas (AIP, F5, IF1 and, in the future, the areas that
constitute the mirror system). Some of them will constitute a matching unit for
the visuomotor correspondences made in AIP. It will transform the simultaneous
inputs from the actual motor program held by AIP motor dominant neurons and
from AIP visual dominant neurons that extract the object’s visual properties to a
common coding, in order to compare them and compute a difference vector, an
operation that is presumably performed by AIP visuomotor dominant neurons,
that will be used to adjust the actual motor program prepared by F5 and
executed by Fl to control the grasping operation. Through an unsupervised
reinforcement learning process the system will tune itself as a neurocontroller
capable of executing complex manipulation tasks, guided by sensorial
information, thus emulating human grasping behaviour. Its real performance
will be tested in our robotic gripper designed in our laboratory using the AER
technique, described later
538

5. The a n t h r o p o m o ~ h~i ~C P T - N ~ ~ O Chand

OR
The biomecha~caldesign of the a n t h r o p o m o ~ ~UPCT-NEUROCOR
c hand
was inspired by the skills of the human hand. It has three fingers and an
opponent thumb, with 4 DOF each of them. They are mounted on a rigid palm.
Each of them has 3 independent joints (MCP, PIP y DIP for the fingers and
CMC, MCP and IP for the thumb). DIP/IP (thumb) and PIPNCP (thumb) are 1-
DOF joints (flexio~extension~ and MCPKMC (thumb) are 2-DOF joints
(flexio~extensionand abductio~adduction).Each DOF is actuated by a pair of
DC Maxon motors and a couple of polystyrene tendons, enrouted through a set
of pulleys through the finger, that stand for artificial muscles, as they emulate
the ago~s~antagonist muscle actuation made by our nervous system

Figure 4. The ~EUROCOR-UPCTanthropomorphic hand. Detail of the fingers, with propioceptive

position Hall effect sensors, tactile FSR sensors and the pulley enrouted tendons that move the
phalanxes in an agonist'antagonist way

Figure 5. The actuation system for the NEUROCOR-UPCT anthropomorphic hand. Each DC motor
(1) moves one polystyrene tendon (2) attached to a horizontal shaft (3) by means of a planetary
gearhead (4). There is an encoder adjoined to each motor to control number of turns.
 539

To measure joint position, speed and direction Hall effect sensors, integrated in
each joint, have been used. Tactile information is read with FSR sensors placed
on the fingertips and on the inner face of medial phalanxes. The hardware
control boards receive commands from a dedicated PC and send backwards the
requested information. through a PCMCIA communication bus. The neural
networks control software is programmed in MatLab.

6. AER in control hardware

NEUROCOR works with other groups in a new research field concerning bio-
inspired hardware systems that exploit the inter-chip communication technique
known as “Address-Event-Representation”(AER). The AER concept, coined in
1991 by Sivilotti [ 181 in Caltech, is being adpoted and developed by a growing
community of researchers interested in the design of VLSI circuits dedicated to
emulate real neural systems (Abusland [ 191).
At first sight, AER is just a TDM asynchronous communication
technique between chips that emulates the high connectivity found between
adjacent layers of neurons by means of a high speed digital bus with few wires.
The idea is to “place” a group of neurons in one chip, another one in another
chip, and use AER to simulate the massive connectivity between both groups.
Moreover, AER is capable of complex operations, as a chip working like an
emitter connected to several receivers or vice versa, such as translations or
rotations that can be made while digital signals are being transmitted.
Among the most ambitious projects regarding AER, we have to
mention the works of IMSE research group, trying to develop microchips
capable of performing real-time convolutions that would have an enormous
potential for artificial vision systems both in a conventional manner (Pitas [20])
and in a bio-inspired way (Sheperd [21]).

7. Conclusion
In conclusion, the innovative aspect of our research work in NEUROCOR
consists in the development of AER-based hardware and bio-inspired neural
networks software to implement a real platform for performing real-time object
manipulation tasks with a robotic gripper in a human-like manner, merging
visual, tactile and propioceptive information in the same control architecture and
controlling the planning and execution of the hand movements emulating the
performance of the central nervous system.

References
1. C. L. MacKenzie and T. Iberall, Advances in Psychology: The Grasping
Hand, 104 (1994).
540

2. M.A. Arbib, The handbook ofBrain Theory and Neural Networks, 13, The
MIT Press (2003).
3. D. Marr, Vision, San Francisco: W.H. Freeman.
4. G. Rizzolatti and G. Luppino, Neuron 31,892 (2001).
5. G. Luppino, A. Murata, P. Govoni and M. Matelli, Exp. Brain Res. 128,
181-187 (1999).
6. G. Rizzolatti & G. Luppino, The handbook of Brain Theory and Neural
Networks, 501-504, The MIT Press (2003).
7. Y. Uno. N. Fukumura, R. Suzuki and M. Kawato, Neural Networks 8, 839-
851 (1995).
8. H. Sakata. M. Taira, S. Mine and A. Murata, Exp. Brain Res. 22, 185-198
(1992).
9. M.A. Moussa, IEEE Transactions on Neural Networks 15,629-638 (2004).
10. A.H. Fagg and M.A. Arbib, Neural Networks 11,1277-1303 (1998).
11. G. Rizzolatti and G. Luppino, Neuron 31, 889-901 (2001).
12. J. Molina Vilaplana, A biologically inspired neural architecture for
learning and control of grasping movement in anthropomorphic robotic
platforms, Ph.D. Thesis, Polytechnic University of Cartagena (2006).
13. J. Molina Vilaplana, A biologically inspired neural architecture for
learning and control of grasping movement in anthropomorphic robotic
platforms, Ph.D. Thesis, Polytechc University of Cartagena (2006).
14. P. Baraduc. E. Guigon and Y. Burnod, Cerebral Cortex 11, 906-917
(2001).
15. A. Murata, L. Fadiga, L. Fogassi, V. Gallese, V. Raos and G. Rizzolatti,
Journal of Neurophysiology 78,2226-2230 (1997).
16. J. Lbpez Coronado, J.L. Pedreiio Molina, A. Guerrero Gonzalez and P.
Gorce, Robbtica 20, 23-3 1 (2002).
17. D. Taddeucci, P. Gorce, Y. Burnod, J. Lbpez Coronado, J.L. Pedreiio
Molina, A. Guerrero Gonzllez, C. Laschi and P. Dado, First IEEE-RAS
International Conference on Humanoid Robots, The Massachusetts
Institute of Technology, Cambridge. MA (2000)
18. M. Sivilotti, Wiring Considerations in analog VLSI Systems with
Application to Field-Programmable Networks, Ph.D. Thesis, California
Institute of Technology, Pasadena CA (1991).
19. A. Abusland, T.S. Lande and M. Hovin, IEEE International Symposium on
Circuits and Systems, ISCAS’96 v01.111,401-404 (1996).
20. I. Pitas, Digital Image Processing Algorithms and Applications, John Wiley
and Sons (2000).
21. G. M. Sheperd, The Synaptic Organization ofthe Brain, Oxford University
Press, 3rd Edition (1990).
 A CONCURRENT PLANNING ALGORITHM FOR DUAL-ARM
SYSTEMS *

JEN-HUI CHUANG AND TING-WE1 CHAN

Depart. of Computer Science, National Chiao Tung University
I001 Ta Hsueh Road, Hsinchu, Taiwan 30010, R.O.C.

CHIEN-CHOU LI"
Depart. of Computer Science and Information Engineering, Shu-Te University
No. 59, Hunshan Rd., Yenchau, Kaohsiung, Taiwan 824, R.O.C.

An algorithm for path planning of object held by multiple manipulators is presented in

this paper. The proposed path planning algorithm is composed with two planners: path
planner of the object and motion planner of manipulators. Both planners use a
generalized potential model to evaluate the repulsion between object/manipulators and
obstacles, so as to avoid collision. By adding virtual obstacle planes to roughly divide the
workspace into several subspaces, a concurrent approach of path planning is also
proposed. With each subspace designated to a manipulator, collisions between
manipulators can be avoided. Therefore, paths of all manipulators can be planned at the
same time.

1. Introduction
While using only one robot in a workspace limits the classes of tasks that
can be performed, more and more multi-robot systems have been considered in
recent decades [ 1..6]. For many applications, manipulators are holding an object
and, together, form a closed linkage, named closed kinematic chains. When
planning the manipulator motion and the object path, the manipulators should be
connected to the object so that the closure constraint is satisfied. Thus, the
complexity of motion and path planning for multi-manipulator systems
increases significantly from single manipulator systems. In [ 13, the motion and
path planning algorithms of closed kinematic chain manipulator systems are
divided into three basic approaches: exact motion planning [ 1, 21, probabilistic
motion planning [3, 41 and mechanism singularity analysis.
This paper proposes a concurrent planning algorithm for chain robot
systems. The proposed algorithm is composed of two stages of path planning

* This work is supported by National Science Council of Taiwan under grant no. NSC 94-2213-E-
366-01 1 & NSC 95-2221-E-366 -010.
Corresponding author, email: jacoblin@mail.stu.edu.tw.

541
542

algorithms: (i) global planning of the object, and (ii) motion planning of the
manipulators. Both algorithms use the generalized potential model [7] to plan
collision-free paths. While (i) is for planning the safest path of the object, (ii) is
for planning safest configurations of manipulators whose end-effectors are
attached to the object.
The remainder of t h s paper is organized as follows. In Section 2, we
propose the concurrent planning algorithm with some implementation details.
Simulations considering more realistic situations are presented in Section 3 to
show the effectiveness of the proposed algorithm. Section 4 concludes this
paper and outlines some possible directions for future works.
2. Concurrently Planning Algorithm
The scenario of this work involves two manipulators trying to move an
object to its destination while avoiding obstacles (including other manipulators).
Since the end-effectors of both manipulators will be attached to the object, the
path planning of manipulators is thus reduced to simpler motion planning with
known end-effector trajectories. Therefore, the proposed algorithm is composed
of two planners: (i) the global planner for the moving object and (ii) the
manipulator motion planner. Similar to the path planning of a rigid object, the
former will find a collision-free path for the object from its start to its goal
location, while the latter is to plan a sequence of manipulator configurations to
hold the object along the planned object path. The generalized potential model
(reviewed next) is used in both planners to avoid collision and to adjust
manipulator configurations for minimum potential (minimum risk of collision).
The generalized potential field proposed in [7] is used to model the 3-D
environment by representing objecdmanipulators by charged sample points, and
obstacles by charged polygons, respectively. In [7] the potential function is
assumed to be inversely proportional to the distance between two point charges
to the power of an integer (m).In particular, it is shown that the repulsive force
exerted on a point charge due to a 3-D polygon can be obtained analytically by
evaluating the gradient of the potential function (for m=3)

= c[&:,Y:,z)- &, +I).

I
y;,
a
-
Z
with (see [ 101 for details)
1 xz , (2)
+(x, y,z) = -tan-'
z y J m
where the xi-y'-zcoordinate system is determined by the right-hand rule for each
edge i of the polygon such that z is measured along the normal direction of S,
 543

and xi is measured along edge i of the polygon, respectively, and a is a constant.

The discussion for repulsive torque is similar and is omitted for brevity.
2.1. Global Path planning
For a rough description of the object path, the proposed approach uses one
or more guide planes (GPs) [8] as final or intermediate goals in the 3-D
workspace. Similar to the approaches in [8], the GPs provide the object a
general direction to move forward. A collision-free traversal of a given
sequence of GPs by the object is regarded as a global solution of the path
planning problem of cooperative manipulators. In this paper, it is assumed the
sequence of GPs is given in advance.
Assuming that a guide plane GPi is given as an intermediate goal, the basic
path planning procedure for moving the object onto GPi is described below:
Algorithm of Global Path Planning
Step 1 Translate the object with distance 6i along the direction of attraction of
the GPi. Find the smallest such that corresponds to a feasible and
collision-free translation. If the object can’t move anymore, go to Step
5; otherwise, add an intermediate guild plane perpendicular to the
direction of attraction.
Step 2 Slide the object without changing its orientation on the GP to minimize
the potential. Considering the forces exerted on the object, let F be the
repulsive force exerted on the object due to the repulsion between the
object and the obstacles. To determine the direction in which the
object should slide on GP, and thus to reduce the repulsive potential,
the projection of the resultant force exerted on the object is calculated.
Step 3 Adjust the object orientation for the minimum potential configuration
with the fixed position. In Step 3, the object is adjusted in orientation
while fixed in position. The direction in which the object should rotate
is determined by the repulsive torque experienced by the object with
respect to its centroid.
Step 4 If the object reaches GP,, the planning is completed. Otherwise, go to
Step 1.
Step 5 Exit and report that goal is unreachable.
For path planning involving multiple GPs, the above algorithm will be
executed for each of them sequentially. It is assumed that the planning for a GP
starts as the planning of the previous GP is accomplished. The path planning
ends as the object reach the goal, which is usually a (goal) GP in the path
planning problems.
2.2. Motion Planning Algorithm of Manipulators
The planning of manipulators can simply be regarded as to finding the
safest configuration for each robot with fixed-position end-effector and base.
The proposed concurrently planning algorithm cuts a chain manipulator into two
544

manipulators and plans two paths of manipulators independently in two

separated subspaces, which are formed by auxiliary obstacle planes.
2.3. Free Space Decomposition
As shown in Fig. l(a), there are five auxiliary obstacle planes to roughly
separate the free space into two subspaces. The sizes, orientations and positions
of these five auxiliary planes is depending on the direction of the object. To
make the explanation more clear, the x-y-z coordinate system of the scene is
depicted as Fig. l(a). The first auxiliary plane, midpoly shown in Fig. l(b), is
formed at the middle of the object. And, other four planes, the sidepolyl,
sidepoly2, sidepoly3 and sidepoly4, are formed above, below, and on two sides
of the object, respectively. While midpoly is attached to the object, sidepolyl
and sidepoly2 are parallel to x-z plane. Fig. 3(c) shows the different orientations
of object through rotation with respect to the z-axis, and with various sizes of
sidepolyl and sidepoly2. sidepolyl and sidepoly2 are in fact the extensions of
the ends of midpoly to the walls of workspace. It is easy to see that the
workspace is thus roughly divided into two subspaces.
Similar to sidepolyl and sidepoly2, sidepoly3 and sidepoly4 are parallel to
y-z plane, and their sizes are also adjustable when the object rotates respective to
z-axis. Similarly, sidepoly3 and sidepoly4 are the extensions of the ends of
midpoly to the walls of workspace.

(a) *I (C)
Fig. 1 (a) A dual-manipulator system has two 7-link manipulators, Master and Slave, their
goals. (b) Five auxiliary obstacle planes. (c) An example of the space decomposition where the
object rotates an arbitrary angle and the positions of the five auxiliary obstacle planes vary as well.

Fig. 2 The forces exerted on distal link and other links of a manipulator with its end-effector
fixed at p of the object.
 545

2.4. Configuration Optimization

As shown in Fig. 2, the direction to which the distal link should rotate is
determined by the total repulsive torque experienced by the distal link with
pivot p. Letfr be the repulsive force from obstacles exerted on J,J (n=3 in Fig.
2). The resultant torque experienced by Ink,, with respect t o p is equal to
r* = T, + f 2 ~ .I, (3)
where I,, is the length of Ink, and frL is the projection of f2 onto a plane
perpendicular to Ink,. To find the minimum potential orientation of Ink,, binary
searches are performed using the projection of T * ~along three orthogonal axes
of rotation, one axis at a time. After the optimal orientation of Ink,, is found, the
orientations of the rest links are adjusted recursively for connectivity and for
minimum potential using T*,,-~, ~ * " - ....,
2 etc.
3. Simulation Results
In this section, we will carry out computer simulation of path planning
algorithms regarding multiple robot arms holding an object. The simulating
platform is based on Windows XP running on a Pentium4 2.8Ghz, 512MB
RAM PC.
In this example we place a narrow passage between two closed spaces. The
object is in the right space initially. We hope to move it to the left space by two
five-link manipulators, as shown in Fig. 3(a). Two GPs (botl and bot2) are used
in this example. We first use botl to guide the object and the robot arms go
through the narrow passage, then use bot2 to pull them to the goal position. Fig.
3(b) and Fig. 3(c) show the partial trajectories of the dual-arm systems.

(4
(a) (C)
Fig. 3 An example within the narrow passage. (a) Initial configuration. (b) The trajectory of
robots from start to botl. (c) The trajectory of robots from botl to bot2.

4. Conclusions
The main purpose of this paper is to route the path for multiple robot arms
holding an object in a three dimensional workspace. The proposed algorithm is
based upon the generalized potential field model which produces repulsive force
between obstacles and the robot arms along with the object, assuring that the
arms and the object will not collide with the obstacles when moving along the
546

planned path in the work space. The concept of virtual spacing plane is
introduced to appropriately divide the workspace into subspaces, each for a
robot arm, in order to resolve the issue regarding robot arms colliding each other
and to make parallel routing possible instead of establishing a process queue for
multiple robot arms. According to computer simulation, the paths we obtain are
moderately smooth and continuous.
In our frame of reference, we only bring out simulations and tests aiming at
the applications of multiple robot arms but did not take in some of the actual
mechanical concerns of robot arm operations such as acceleration, energy
consumption, and the load of the arms and joints. Nonetheless, the path we
obtained is still valuable as a start or a reference for further detail modifications
for actual applications. Our routing solutions then become a bridge between
theory and reality.
References
1. J. C. Trinkle and R. J. Milgram, “Complete path planning for closed
kinematic chains with spherical joints,” Intl. Journal of Robotics Research,
vol. 21, no. 9, pp. 773-789, Sep. 2002.
2. G. F. Liu and 3. C. Trinkle, “Complete path planning for planar closed
chains among point obstacles,” in Proc. Robotics: Science and Systems,
2005.
3. J. H. Yakey, S. M. LaValle, and L. E. Kavraki, “Randomized path planning
for linkages with closed kinematic chains,” IEEE Trans. on Robotics and
Automation, vol. 17, no. 6, pp. 951-958, 2001.
4. J. Cortts, T. Simton, and J. P. Laumond, “A randomized loop generator for
planning the motions of closed krnematic chains using PRM methods,” in
Proc. IEEE Intl. conf. Robotics and Automation, vol. 2, pp. 2141-2146,
2002.
5. K.-S. Hwang, M.-Y. Ju and Y.-J. Chen, “Speed Alteration Strategy for
Multijoint Robots in Co-Working Environment,” IEEE Trans. on Industrial
Electronics, vol. 50, no. 2, pp. 385- 393, 2003.
6. P. Tournassoud, “A strategy for obstacle avoidance and its application to
multi-robot systems,” Proc. of the IEEE Int. ConJ on Rob. & Auto. (ICRA),
pp. 1224--1229, 1986.
7. J-H. Chuang, “Potential-based modeling of three-dimensional workspace of
the obstacle avoidance,” IEEE Trans. on Rob. & Auto., vol. 14, no. 5, pp.
778-785, 1998.
8. C.-C. Lin, J.-H. ’ Chuang, “Potential-based path planning for robot
manipulators in 3-D workspace,” IEEE Int. Conf. on Rob. & Auto. (ICRA),
pp. 3353-3358, 2003.
 A modular approach for controlling mobile robots

K. Regenstein, T. Kerscher, C. Birkenhofer T. Asfour, J..M. Zollner, R. Dillmann

Interactive Diagnosis- and Servicesystems (IDS), Research Center f o r Information
Technologies (FZI),
Karlsruhe, 76131, Germany
* E-mail: regenstein@fzi.de, kerscher@fzi.de, birkenhofer@fzi.de, asfour@fzi.de,
zoellner@fzi. de, dillmann@fzi.de
www.fzi.de/ids

A variety of different robots was built at our institute. As these robots differ as
well in size, shape and in actuation principle it would be very time consuming
and inefficient to tailor a computer and hardware architecture especially to the
specific robot. In this paper it will be described how common aspects in robot
control can be identified and how modular hardware components can be derived
from a modular software framework and a respective computer architecture. A
decentralized computer architecture based on embedded PC systems connected
to local controller modules via CAN Bus was developed. The requirements and
restrictions that led to the development of these controller modules and their
associated power amplifier boards will be described

Keywords: Computer Architecture; Modular Control Concept; Hard-

ware/Software Co-Design

1. Introduction
In a large number of robotic systems a decentralized architecture is
At the Research Center for Information Technologies (FZI) different kind
of robots - like humanoid robots, four- or six-legged walking machines,
mobile platforms and snakelike sewer inspection robots - are developed.
For these robots we designed a computer architecture based on embedded
PCs and distributed controller modules connected with each other via one
or more CAN-BUSS~S.~ Though the requirements concerning the distributed
components are quite different we wanted to implement a persistent design
that could be used in all robots with only small amount of adaptation.
The main issues for the controller modules used in our robots are space
requirement, power consumption, several inputs for sensor value acquisition
and communication interfaces (i.e. CAN-Bus). As none of the available of-

547
548

the-shelf products suited all these needs we decided to build a controller

module and an associated power amplifier ourselves.

2. Computer Architecture
The mechatronical construction of a robot can roughly be divided into me-
chanical aspects and into aspects of setting up the electronic and computer
system. In this section we will describe how the electronic system of our
robots is set up and how a computer architecture suiting the needs in these
robotic systems was designed.
We started by identifying the concepts of how a robot should accomplish
given tasks and thus proposing a control architecture that then was the
basis for developing the computer architecture. We chose a hierarchically
organized control system for the robots with the three following levels:6
0 The task planning level specifies the subtasks for the multiple sub-
systems of the robot. Those could be derived from the task descrip-
tion autonomously or interactively by a human operator
The task coordination level generates in sequence/parallel primi-
tive actions for the execution level in order to achieve the given
task goal. The subtasks are established by the task planning level.
The execution of the subtasks in an appropriate schedule can be
modified/reorganized by an operator using an interactive user in-
ter face
0 The task execution level is characterized by control theory to ex-
ecute specified sensory-motor control commands. This level uses
task specific local models of the environment and objects, which
represent the active scene
According to the control architecture the computer architecture is struc-
tured into three levels as well. Choosing suitable devices for these three lev-
els yielded that the requirements of the task planning and task coordination
level could be met with industrial PCs and PC/104 systems. As cabling in
the robot is a major issue it is desirable to reduce cabling efforts as much
as possible. Because of this we decided to use a decentralized system for
sensory-motor control. By placing controller modules close to the motors
and sensors cabling can be reduced to one common power supply and a bus
connection. Wires for supplying the motors and connecting the sensors to
the controller can be kept short and have not to be passed through moving
joints. To fulfil the requirements of the task execution level we designed the
so called Universal Controller Module (UCoM) that in combination with our
 549

Fig. 1. The Universal Controller Module (UCoM) (upper left) and the 3way-brushdriver
(lower left); Schematic overview of data flow on the UCoM and to the piggyback board

motor controller Sway-brushdriver (Fig. 1) is responsible for the sensory-

motor control of the robot. The features of the combination of UCoM and
3way-brushdriver will be described in section 3 in more detail.
As software framework we use the Modular Controller Architecture
(MCA2)7 that was developed at our institute and is available under GPL
online here.8 The idea behind MCA2 is to structure the software into
reusable modules with simple interfaces. Each module has the three data
channels: control data, sensor data and parameters. Via these data channels
information is exchanged between the modules. A number of modules can
be combined into a group which has the same interface as one module.

3. Controller modules on sensory-motor level

Following the modular strategy that is realised in the software framework
MCA2 we wanted to reach this modularity in the computer architecture
as well. To achieve this goal not only for one of our robots but spanning
all different robots we have to choose a common controller unit. This is
important to reduce programming efforts as well. For example you can
implement a PID-controller only once and as you use the same hardware
that can run the same software in different robots you only have to adapt
the PID parameters for the chosen joint. As already mentioned the main
issues for the controller modules are space requirement, power consumption,
several inputs for sensor value acquisition and co~municationinterface.
Especially for motor control the mandatory requirements for the controller
module were
5 Suitable to control three brushed DC motors at 24 V at up to 5 A
5 Achieve cycle times as low as 1ms
550

0 Able to decode six quadrature coded signals

0 Small outline, positioning close to the actuator possible
0 Low power consumption
0 Interface to access CAN-Bus

In some applications the 5 A might not be sufficient but this was the maxi-
mum current that could be realized without exceeding the space limitation.
Besides electrical motors we use other actuation principles in our robots.
For example in one of our walking machines - Airbugg - fluidic muscles
were used and there is still ongoing research evaluating fluidic muscles as
actuation. For this kind of actuation a valve driver is needed. Furthermore
in some robots like LAURON’O we need extended sensor input like posture
information from gyroscopes and acceleration sensors. To avoid building a
special controller module for each of these applications we decided to split
the controller module into one part that actually contains the controller and
one part that contains the power amplifier, the valve driver or sensor acqui-
sition electronics. As mentioned above we named the part with the actual
controller Universal Controller Module (UCoM) as it will be universally
used in our robots together with the respective piggyback board.

3.1. Universal Controller Module - UCoM

The choice for a suitable microcontroller/DSP for the UCoM could be nar-
rowed down rather quickly as nearly no available controller had all the
required features. The Freescale ”DSP56F803 16-bit Hybrid Controller” l 1
came closest to our needs. This controller is a DSP featuring a set of pe-
ripherals usually only known from microcontrollers.
Though this hybrid controller nearly matches the requirements it still
misses some essential features. On the one hand side it does not have enough
general purpose 10s to control three motors on the other hand it only has
two quadrature timers capable of decoding quadrature coded signals. To
extend the DSP’s flexibility we decided to put an FPGA next to it. As
suitable FPGA we chose the Altera EPFlOk30A. With this FPGA we can
equip the UCoM with a high number of general purpose 10s. We gain a
high flexibility concerning routing and assignment of pins to the piggyback
board. Through the FPGA we can reassign most of the signals so that it
suits the used .piggyback board. The FPGA is also used to preprocess data
that is exchanged between the UCoM and the piggyback board.
By this approach we can disburden the DSP from tasks that are done
in hardware more efficiently. For example we implemented six decoders for
 551

quadrature encoded signals. The communication between DSP and FPGA

takes place via the external memory interface. As FPGA and the exter-
nal RAM that we integrated on the UCoM share the external address
range we implemented an address decoder into the FPGA. This address
decoder deasserts the chipselect for the lowest 64 addresses of the external
address range and receives the sent data. For all other addresses the data
is routed t o the external RAM so that nearly the whole external memory
range is available to the DSP and only the lowest 64 addresses are used as
FPGA-registers. The DSP always initiates communication with the FPGA
by writing to or reading from a FPGA-register. These registers are used to
exchange data between the two devices. So t o get the value of a quadrature
coded signal all the DSP has to do is access the respective external RAM
address.
The UCoM combines the Freescale "DSP56F803 16-bit Hybrid Con-
troller", an external RAM and an Altera EPFlOk30A on one board. It
is interfaced to the desired piggyback board via two 60-pin 0.8mm pitch
board-to-board connectors. Via this connector the UCoM is supplied with
a 5 V power supply. From this 5 V we generate the 3.3 V that are needed on
the UCoM. We do not directly feed 3.3V t o the UCoM t o avoid problems
with voltage drop or disturbances that must be expected due t o the wiring
close to the motor power wires. The 3.3 V generated on the UCoM are then
fed back to the board-to-board connector to be available on the piggyback
board.

3.2. Motorcontml Board

As the main actuation principle in our robots are electronic motors the
first piggyback board we developed was a power amplifier able t o drive three
brushed DC motors. We named this piggyback board Sway-brushdriver. On
the 3way-brushdriver we integrated three H-Bridges which are driven by a
3-phase brushless DC motor controller chip each. This motor controller chip
can be configured to drive brushless or brushed motors. To drive brushless
motors the hall-inputs to the driver must be connected to the hall-sensors
of the motor. To drive brushed DC motors the hall-inputs are simply tied
t o ground. We chose this motor controller chip so that it can be interfaced
by software in the same manner if we need t o design a piggyback board for
brushless DC motors in the future. In both branches of the H-Bridge we
integrated a shunt via which we can measure the motor current for each
motor. As the motors are driven by PWM signals we had t o use an OP-
AMP with a high gain-bandwidth-product to amplify the signals for use in
552

Fig. 2. Robots in which the presented computer architecture is actually used: AR-
MAR 111 (left) and LAURON IV (right)

the AD-Converter.
As we wanted to keep the UCoM as small as possible we decided to put
all interface connectors except the ones that are directly wired to the DSP
like CAN-Bus, serial communication interface and JTAG to the piggyback
board. Thus the piggyback board is responsible for supplying the UCoM
with the 5 V input voltage. The 3way-brushdriver has a connector for in-
put of the aforementioned 5V, 24V as power supply t o the motors and a
common Ground.
Further connectors on the motor control board are six small connectors
for the quadrature coded encoder signals. Each of these connectors has six
pins, two of which carry supply voltage of 5 V and ground for the encoder,
two are the quadrature channels A and B. The remaining two are an index
signal and a pseudo absolute code. A schematic overview of the dataflow
on the UCoM and to the piggyback board can be found in Fig. 1.

3.3. Software C o ~ p o n e n t son DSP

To extend the modularity down t o the code for the DSP each UCoM is
treated as one module in the MCA2 software framework. Thus the UCoMs
can be seamlessly integrated into the software running on the linux PC via
a MCA-driver interfacing to the CAN-Bus. A number of different basic ap-
plication programs for the UCoM were developed so that they can be used
in most of the robots with only small adaptations. In the humanoid robot
ARMAR-I1Il2 we only use one generic program for all different joints. This
generic program can be adapted via a configuration file that is evaluated
at system startup. Other basic application programs that are implemented
are: direct-pwm, a p-controller, a pid-controller and a SpeedlPosition Con-
 553

troller. Further control programs including for example time discrete and
torque control algorithms are in development.
To download these programs t o the UCoM a bootloader that accepts
data via CAN-Bus is used. This is very convenient if programming is done
frequently in the development phase.

3.4. Functions on FPGA

The FPGA can be seen as an extension to the DSP. It equips the DSP with a
large number of general purpose 10s and processes data that is exchanged
between the DSP and the piggyback board. As stated above we use the
memory interface to communicate between DSP and FPGA. The function
blocks in the FPGA are programmed in VHDL. There are some blocks that
are common t o all designs independent of the plug on board: An Address
Decoder, a version supervision sytem, Initialization and a Watchdog.
For the introduced Sway-brushdriver the following function blocks were
already implemented into the FPGA: A Quadrature decoder, Pseudo Ab-
solute Decoder, Serial Synchronous Interface and a Motor Control register.
Further modules easily can be integrated into the FPGA as they are needed.

4. Summary and Outlook

In this paper we presented a modular concept to control robots which we
applied to our robots ARMAR I11 and LAURON IV (depicted in Fig. 2).
This concept includes a control architecture from which the used computer
architecture was derived, a modular software framework and the develop-
ment of a hardware architecture that can be mapped into the computer
architecture. The focus of this paper was on setting up a modular system
that can be used in a variety of robots so that not only software components
can be reused but that also the hardware is interchangeable. It was laid out
how this goal was achieved and especially the development of the UCoM
and the 3way-brushdriver were described. Some examples of our robots in
which these control devices are successfully used were presented above. In
ongoing research we will implement this concept in robots that are going
to be built. Most likely further piggyback boards will be designed for that
purpose.

5. Acknowledgement
This work is partially supported by the German Research Foundation
(DFG) within the Collaborative Research Centre SFB 588 (Humanoid
554

Robots - Learning and Cooperating Multimodal Robots).

References
1. J.-Y. Kim, I.-W. Park, J. Lee, M.-S. Kim, B. kyu Cho, and J.-H. Oh, “System
Design and Dynamic Walking of Humanoid Robot KHR-2,” in Proceedings
of the 2005 IEEE International Conference on Robotics and Automation,
Barcelona, Spain, April 2005, 18-22 April 2005, pp. 1431-1436.
2. T . Sugihara, K. Yamamoto, and Y. Nakamura, “Architectural design of
miniature anthropomorphic robots towards high-mobility,” in Intelligent
Robots and Systems, 2005. (IROS 2005). 2005 IEEE/RSJ International Con-
ference on, 2-6 Aug. 2005, pp. 2869-2874.
3. J. Butterfass, G. Hirzinger, S. Knoch, and H. Liu, “DLR’s multisensory ar-
ticulated hand. I. Hard- and software architecture,” in Robotics and Automa-
tion, 1998. Proceedings. 1998 IEEE International Conference on, vol. 3, 16-20
May 1998, pp. 2081-2086~01.3.
4. R. Bischoff and V. Graefe, “HERMES - a versatile personal robotic assistant,”
Proceedings of the IEEE, vol. 92, no. 11, pp. 1759-1779, Nov. 2004.
5. K. Regenstein and R. Dillmann, “Design of an open hardware architecture for
the humanoid robot ARMAR,” in Humanoids 2003, International Confer-
ence on Humanoid Robots, Conference Documentation, October 1 - 3, 2003,
Karlsruhe and Munich, Germany, October 2003, p. 3.
6. T. Asfour, D. Ly, K. Regenstein, and R. Dillmann, “Coordinated Task Exe-
cution for Humanoid Robots,” in Experimental Robotics IX. Springer Berlin
/ Heidelberg, 2006, vol. 21, pp. 259-267.
7. K. U. Scholl, J . Albiez, and B. Gassmann, “MCA - An Expandable Modular
Controller Architecture,” in 3rd Real-Time Linux Workshop, 2001, Milano,
Italy, 2001.
8. FZI, “Modular Controller Architecture Version 2.” [Online]. Available:
http://www.mca2.org/
9. K. Berns, J. Albiez, V. Kepplin, and C. Hillenbrand, “Airbug - Insect-like
MachineActuated By Fluidic Muscle,” in Proceedings of CLAWAR 2001, Int.
Conference on Climbing and Walking Robots, 2001.
10. B. Gassmann, T . Bar, J. Zollner, and R. Dillmann, “Navigation of Walking
Robots: Adaptation to the Terrain,” in Proceedings of CLAWAR 2006, 9th
International Conference on Climbing and Walking Robots (CLA W A R ) , 2006.
11. Freescale, DSP56F803 Data Sheet. [Online]. Available: http://www.freescale.
com/files/dsp/doc/datasheet/DSP56F803.pdf
12. T. Asfour, K. Regenstein, P. Azad, J. Schroder, A. Bierbaum,
N. Vahrenkamp, and R. Dillmann, “ARMAR-111: An Integrated Humanoid
Plattfrom for Sensory-Motor Control,” in Proceedings of IEEE-RAS Inter-
national Conference on Humanoid Robots, 2006.
 A SELF ORGANIZING NETWORK MODEL FOR CLAWAR
SYSTEMS COMMUNICATION COEVOLUTION

FABIO P. BONSIGNORIO
RTD, Heron Robots s.r.l., V.R.Ceccardi, 1/18
Genova, I-16121, Itah

This paper shows a model of robot-robot communication that is based on the concept of
'network embodied cognition'. The communication between intelligent clawar agents is
seen as an activity made possible by the self-organisation of coevolving 'situated' and
'embodied' cognitive processes, physically distributed among the inter-communicating
agents, motivated and initiated by physical 'finalized' interactions with the environment.
The benefits and trade-offs of this approach for managing swarms of CLAWAR systems
are compared and discussed against alternative solutions.

1. Introduction
This paper exposes a model for the evolution of a communication pragmatics
between networks of CLAWAR autonomous agents, between them and in
perspective with humans.
In nature there are many h d s of loosely coupled networks of intelligent agents,
largely varying in terms of quantity of agents and cognitive and adaptive capacity
(i.e. of computational needs) of each agent.
In the natural domain the most widely used method of 'intelligence', computation
and 'cognition' are 'embodied' biological neural networks.
A number of empirical and theoretical researches, [2,3,4] are investigating, on
one side on the aspects and implication of 'embodunent', particularly interesting
in the walking machine domain,[l5], on the other side on the 'emergence' of
cognition from network interaction of physical agents, [7,8,9].
We propose here a model of cognition were the model of the environment
'emerges' from the collective interaction of a networked robot system with its
environment.
We assume (for simplicity) that the 'cognitive network ' can be accessed by all
the agents which were/are coevolving it and in fact share (constitute) it.
In this conceptual fiamework 'models of the environment' are shared very
naturally, while is needed a concept of 'self.

555
556

This is actually local to the agent and evolved by means of evaluation of the
interaction of the specific agent with the environment.
By means of random action generation the nodes which have no consequences on
the individual agent are labeled by that agent as b o n self.
This mimic the basic behavior of the human and mammals immune system (which
can be regarded by itself as a 'cognitive system'.

2. The cognition model

The self organizing coevolutive model of communication, and cognition,

proposed here, consider the building of the model of the environment by a
number of networked physical agents, a 'swarm' , 'situated' in a an evironment
and interacting with it, as a collective learning process represented by the growth
of a network of nodes mapping, for simplicity, the already identified sensory
motor correlations.

The random activity of the network of moving and interacting CLAWAR agents
allow the system to identify the regular patterns in the variables (sensors and
actuator propioceptors data flow) and to connect them in a model of the
environment.

After that the first node has been created, new nodes (highly correlated groups of
sensors and actuators) are attached preferentially to the previous one, according
to their 'fitness'.

The resulting representation is shared (again as a simplification)between all the

agents.

In the following paragraph some information are given about the modeling of
network growth process, the fitness function wluch drive the preferential
attachment of new nodes and the 'self model of each agent.

2.1. Networkgrowth model

In this section we review the recent fast progress in statistical physics of
evolving networks, see [21], with the aim of groundmg a model of
communication for physical automous agent typically represented by CLAWAR.

So far researcher interest has focused mainly on the structural properties of

random complex networks in communications, biology, social sciences and
 557

economics.
A number of giant artificial networks of such a kind came into existence recently.
This opens a wide field for the study of their topology, evolution, and complex
processes occurring in them.
Such networks possess a rich set of scaling properties. A number of them are
scale-free and show striking resilience against random breakdowns. In spite of
large sizes of these networks, the &stances between most their vertices are short,
a feature known as the “small world” effect.
It is known that growing networks self-organize into scale-free structures
through the mechanism of preferential linking.
Numerous networks, e.g.; collaboration networks, public relations nets, citations
of scientific papers, some industrial networks, transportation networks, nets of
relations between enterprises and agents in financial markets, telephone call
graphs, many biological networks, food and ecological webs, metabolic networks
in cell etc., can be modeled in this way.
It is thought that evolving self organizing networks can (could) model the
collective knowledge of a network of intelligent (artificial) autonomous agents.
An ‘objectJprocess’ in the environment is modeled by a node of the network
with many links generated by a fitness process within a coevoluted learning
process.
In our model the probability that a new node will connect to a node i already
present in the network is a function of the connectivity ki and on the fitness qi of
that node, such that:

A node i will increase its connectivity ki at a rate that is proportional to the

probability that a new node will attach to it, giving:

The factor m accounts for the fact that each new node adds m links to the
system.
Thus the connectivity distribution, i.e. the probability that a node has k links, is
558

given by the integral:

We define a proper Q function which is basically a performance index of the

effectiveness of sensory motor coordination.

2.2. Thefitness function

In our model the fitness function of the node i, qi ,controls the growth process of
the cognitive network.
It is given by a function of the mutual information, between the sensors and the
actuators connected to that node. The mutual information between two given
variables is given by equation (4), where X and Y two random variables:

If X and Y are statistically independent eq. 4 gives I(X,Y)=O

It is possible that a more sophisticated performance index, including for instance
'transfer entropy' [ 101 could be helpful, as mutual information doesn ' t give any
indication about the direction of information flow from X to Y, or the opposite,
but only a measure of correlation between the two variables.
It is interesting to notice that entropy based performance indexes have been
proposed to guide the search within evolutionary algorithm as in [ 1 I ]

It is important to notice that nodes are not coincidmg with the physical
CLAWAR agents.

They can be sub processes of a single robot system, for instance the coordination
process between the vision sensors and the front legs, or the one between the
infrared and the sound sensors.

2.3. The 'self 'model

As told, in the model of cognition described in this paper the sharing of the
model of the environment is very natural, while of the contrary, a model of 'self'
 559

must be defined explicitly.

Each agent is able distinguish its 'self from the 'environment' with a parallel
selection process inspired by the mammals immune system.
The key concept is that any physical agent will activate at random all the network
node actuators and will identify which of them will trigger a change in the
environment measuring the transfer entropy between the actuators of the node
and the related sensors.
The 'self of the robot will be the list of nodes for which it has experimented
consequences in the environment.
This is similar to the 'self identification process in some models of the immune
system, [ 121.

3. A toy system representation

A simplified model of what we are discussing is described in this paragraph.
In Table 1 is represent the Cartesian product between the sensor measure
(streams) Si and the actuator propioceptor values Aj,,of all the CLAWAR agents
of the 'swarm'.
The nodes are created when the fitness function ,representing the mutual
information between the nodes, is above a 'relevance' threshold.
The nodes are created as the system evolve in time, as the individual agent
'explore' randomly the environment. Any new node attach preferentially to the
already existing according to their 'fitness I.

Table 1. A simplified represenataion of the node mapping in a example

case.

In figure 1 is represented the resulting connected graph of the nodes.

Table 1 and figure 1 represent the 'model of the world' shared by the networked
robot systems.
560

Figure 1. Node connection graph

In figure 2. is depicted the representation of the 'self of an agent.

Figure 2. Schematic representation of an agent self.

As in the 'real' environment' we should have 'many' sensors and actuators

belonging to many CLAWAR systems the network will grow in accordance to
the equations recalled in paragraph 2.1 .
 561

4. Computing Backbone Architecure

As it should emerge from the previous discussion, the architecture of this
cognition system is inherentlyhighly parallel.
We think that the advantages of this kind of algorithm should be emphasized by
the appropriate distribution of the modules constituting the system over a hghly
parallelized multiserver backbone computing facility.
We describe below how we plan to implement the cognition system on an
advanced grid based distributed architecture, named RoboGrid [22].
The n-tiered structure of the RobotGrid middleware is depicted in Fig. 3. At the
lower layer we have the infrastructure resources, we have the grid services layer,
which implements the virtualization of the resource of the bottom level,
according to the SOA (Service Oriented Architecture) paradigm on demand.
Above we have the application server level, where the (e.g. enterprise java) beans
implementing the planning and control logic reside.
J2EE
Implementation

1 resources 11 services 1 ~ p Sewer

p IWeb server 1,

Figure 3. The n tiers

We show in figure 4 the flow of the sensefreaction scheme of a single robot.

The ‘applet’ running on the robot onboard virtual machine sense e.g. a
movement of a ‘focus object’ in the ‘environment’.
Sense/reactton
scheme
lnrnhec grid ,vc\

Figure 4. The sense/reaction scheme

562

A front end unit of service (an interface 'enterprise java bean') is invoked with a
string of parameters , the bean activate the necessary actions instantiating the
necessary transient grid services (adding resources where necessary) and some
where another bean dispatch a message that trough an on board applet act in the
environment (e.g. moving one or more robot on the field from one position to
another) .

All sensor and actuator information and the network representation of the
environment are shared.
Only the 'self information is stored in the single agent.

5. Discussion
The approach described here has something in common with evolutionary
programming and particle swarm optimization and 'swarm intelligence' in general
as it is an emergent coordination process, on the other end it is 'deeper' as it
generates and 'emerging' model of the environment shared among the networked
robot systems.
This occurs at the price of a certain complexity, or at least, not immediately
intuitive 'simplicity', in the implementation.
In comparison to more traditional centralized planning and control schemes
based on the so called 'symbolic paradigm' , it has the main advantage in the fact
that doesn' t need, in principle, any explicit programming of knowledge of the
environment.
It also seems a good model for 'natural cognition' processes if the model is
extended to take care of 'embodiment'.

6. Futher Research
From an experimental perspective it would be useful to benchmark in a practical
situation against other more classical approaches in order to verify if it could
enable more more robust and autonomous applications of networked CLAWAR.
From a theoretical standpoint a formalization of the 'toy system' in section 3
could help to develop a quantitative framework for 'embodied' and 'emergent'
cognition in robot systems in general and CLAWAR in particular.
An application to the self organization of environment cognition in single robot
systems (including humanoids) seen as 'network' of (semi) independent sensory
motor units could also be, possibly, of some interest.
 563

7. Conclusions
We have described a model of self organizing communication and cognition
schema that allow the development of new architecture networked robot
systems, which could, in principle, show more autonomy than current one, on the
other side this model could help formalize a (semi) quantitative theory of the
emergence of cognition in ‘embodied‘agents.

References
1. C.E. Shannon, “The Mathematical Theory of Communication”,
Bell Sys. Tech. J. 27,379,623 (1948).
2. R.A Brooks, “Robust Layered Control System for A Mobile
Robot” IEEE Journal of Robotics and Automation, (1986)
3. R. Pfeifer, “Cheap designs: exploiting the dynamics of the
system-environment interaction. Three case studies on navigation.”,
In: Conference on Prerational Intelligence --- Phenomonology of
Compexity Emerging in Systems of Agents Interacting Using
Simple Rules. Center for Interdisciplinary Research, University of
Bielefeld, 8 1-9 1, (1993).
4. R. Pfeifer, F. Iida, “Embodied artificial intelligence: Trends and
challenges. Embodied artificial intelligence” Iida et al. (Eds),
LNCS/AI Vol. 3139, 1-26, Springer, (2004).
5. H. Touchette, S. Lloyd, “Information-theoretic approach to the
study of control systems “,Physica A 331, 140-172, (2004).
6. P. Bak, C. Tang, K. Wiesenfeld, “Self-organized Criticality: An
Explanation of l/f Noise” Phys. Rev. Letter. 59,4,381-384, (1987).
7. G. Gomez , M. Lungarella, D. Tarapore, “Information-theoretic
approach to embodied category learning” Proc. of 10th Int. Conf.
on Artificial Life and Robotics, pp.332-337,(2005).
8. D. Philipona, J.K. 0’ Regan, J.-P, Nadal, 0.J.-M.D. Coenen,
“Perception of the structure of the physical world using unknown
multimodal sensors and effectors”, in Advances in NeuraZ
Information Processing Systems, (2004).
9. L. Olsson, C.L. Nehaiv, D. Polani, “Information Trade-offs and
the Evolution of Sensory Layouts” In Proc. Artificial Life IX,
(2004).
10. M.Lungarella, 0. Sporns, “Mapping information flow in
sensorimotor networks”, PLOS Computational Biology,
2/10,1301-13 12, (2006).
11.S.Liu, M.Mernik, B.R. Bryant, “Entropy-driven parameter
control for evolutionary algorithms”, Informatica, 31, 42-50,
(2007).
12. D. Morpurgo, R. Serenitb, P. Seiden, F. Celada, “Modelling
564

thymic functions in a cellular automaton”, International

Immunology, 714, 505-5 16, Oxford, (1995).
13. P. Gacs, “The Boltzmann Entropy and Randomness Tests” In:
Proc. 2nd IEEE Workshop on Physics and Computation
(PhysComp’94), 209-2 16, (1994).
14. P. Gruenwald, P. Vitanyi, “Shannon Information and
Kolmogorov complexity”, IEEE Transactions on Information
Theory, (2004).
15. M. Garcia, A. Chatterjee, A. Ruina, M.Coleman, “The Simplest
W a k n g Model: Stability, Complexity, and Scaling,” Transactions
of the ASME, Journal of Biomechanical Engineering, 120, 281-8,
(1998).
16. S. Lloyd, “Use of mutual information to decrease entropy:
Implication for the second law of thermodynamics”, Phys. Rev. A,
39,10, 5378-5386,(1989).
17. N. Wiener, “Cybernetics: or Control and Communication in the
Animal and the Machine”, MIT Press, Cambridge, MA, (1948).
18. J.C Maxwel1,”Theory of Heat”,Appleton, London, (1871).
19. J. Kennedy , R. Eberhart,”Particle swarm optimization”,Proc.of
the IEEE Intl. Conf. On Neural Network, Washington DC,
USA,1942-1948, ~01.4(1995).
20. M. M Millonas, “Swarms, Phase transitions, and Collective
Intelligence” In C.G. Langton (Ed.), Artificial Life 111, pp. 417-
445, Santa Fe Institute Studies in the Sciences of the Complexity,
Vol. XVII, Addison-Wesley , Reading, Massachusetts,( 1994).
2 1.R.Albert, A.L. Barabasi, “Statistical physics of complex
networks”, Rev. Mod. Phys. 74,47-97,(2002).
22. F.P. Bonsignorio, “A grid based distributed multiagent enabling
system for intelligent autonomous robot swarms”, ISR 2005,
http://www.isr2005.com/,Tokyo, (2005).
23.F.P. Bonsignorio, “Preliminary Considerations for a Quantitative
Theory of Networked Embodied Intelligence”, in Special issue
AI50, LNAI, Springer, Heidelberg, (2007) (to appear)
 AN APPROACH TO GLOBAL LOCALIZATION PROBLEM
USING MEAN SHIFT ALGORITHM

GIOVANNI MUSCATO, SALVATORE SESSA

D.I.E.E.S. Universiv ofcatania, Viale A . Doria 6
Catania 97100, ITALY
vmztscuto(u)diees.unicl.it, sse.ssa@,diees.unict.it

This paper describes a global localization algorithm (GLA) for a mobile robot in indoor
environments, based on a particle filter. This algorithm uses data obtained by two kinds
of sensors: encoder and scanner laser. Given a map of the environment, where the robot
moves, a GLA tries to localize the robot on the map by using its sensor data. The map of
the environment is preliminary built mixing laser readings from well-known poses of the
robot. The mean shift algorithm (MSA) processes the map and obtains a list of features,
which is the synthetic map of the environment used in the GLA. The MSA is also applied
for each sampling step in order to calculate the importance factor of the particles. The
trials have been performed by using a dynamic simulator of a differential drive robot and
the 3Morduc mobile robot.

1. Introduction
Robot localization is the problem of estimating the robot pose relative to a fixed
external reference frame. A Global Localization Algorithm (GLA) tries to
localize the robot on the map by using its sensor data and the map of the
environment. In other words, the GLA answers to the question “Where is the
robot?” placing a robot somewhere in a known environment [ 11. Grid-based
techniques were among the firsts GLAs implemented; the map is divided into
cells and their values represent the probability of the robot to be located in it.
Unfortunately, the computational requirements are proportional to the square of
the dimension of the map. Simmons et al. have applied this method in [2] to
localize a robot in an office environment and Burgard et al. in [3] to localize a
robot in a museum. A solution to these computational problems is the map
matching technique that can be applied directly with the raw data of the sensors,
as described in [4] by Konolidge and Chou, or by identifying suitable landmarks
of the environment, as described by Thrun in [S]. Motivated by the famous
condensation algorithm used in computer vision, Dellaert et al. in [6] and Fox et
al. in [7], were the first to develop a particle filter for mobile robot localization.
Jensfelt et al. in [8] applied Montecarlo localization to features based maps.

565
566

In the implemented algorithm, the map of the environment is a list of

features obtained by using the Mean Shift Algorithm (MSA) [9]. The
implemented algorithm relies on a particle filter; each sample contains the
estimated pose of the robot (position and heading). In the measure phase, when
a scanner laser measure is ready, the MSA computes the local map of the
features. The importance factor of the samples can be updated by evaluating the
distance of the recognized features (in the local map) from their positions in the
global map. The proposed algorithm has been tested by using a dynamic
simulator of a differential drive robot based on Open Dynamic Engine library
[lo]. Moreover several experimental trials have been performed on the
3Morduc, one of the mobile robotic platforms available in our laboratory
(Fig. 1).

Figure 1.3Morduc mobile robotic platform.

In order to increase the interoperability between the simulator and the real
environment, the software architecture of the robot is fully compatible with the
exchanged commands between the GLA application and the simulator software
ill].

2. The clustering algorithm

The MSA is a classical way to extract features from images in computer vision.
The algorithm shown is an adaptation of that described in [9] to scanner laser
measurements with, in additions, some rules to extract clusters. These clusters
represent the features extracted from the environment, used to build the global
map during the initialization phase and the local maps in the execution phase.
The main target of the MSA algorithm is that of finding all the local maxima of
the multivariate kernel density estimate function. That function is obtained as
sum of all the kernel function each centered in a laser scanner measure point and
its local maxima represent all the point in which the scanner laser measurements
 567

are more dense as explained below. The algorithm is iterative and has been
demonstrated in [9] that it is convergent.
Given a scanner reading {xi),=,,,,n,
the multivariate kernel density estimate is
obtained as follows:

Where K(x) is a kernel hnction with windows radius h. The optimum

kernel function yielding minimum mean integrated square error is the
Epanechnikov kernel

I o otherwise
The use of a differentiable kernel allows computing the estimate of the
density gradient as the gradient of the kernel density estimate:

Where Mh(x) is the sample mean shift vector:

Where &(x) is a circumference of radius h centered in x and containing n,

points. The mean shift vector has the direction of the gradient of the density
estimate at x and points towards the direction of the maximum increase in the
density. As a result, the mean shift algorithm is executed by iterating the two
phases: computation of the mean shift vector kth(x) and translation of the
window sh(X) by Mh(X)
The iteration process ends when the components of Mh(x) are less than a
threshold. To improve the independence of the feature determination by the
distance of the scanning point, the h radius is expressed as:

h, = kdi (4)

Where k is the feature compression parameter and influences the final

number of features found for each laser scan. Fig.2 shows the results of the
568

feature extraction algorithm applied to a square room of 4 x 4 meters by using

three different values of k. If k is too small (Fig.2a) all the scanning points are
evaluated as features, on the contrary, if k is too big (Fig.2~)the algorithm will
found only one features in the barycenter of the scan. For this reason, the tuning
of this parameter is fundamental to obtain good feature extraction results
(Fig.2b). After this pre-elaboration phase, the points (the n centers of the
windows) are clustered by considering their distances. The barycenter of the
cluster is the representative point of the set, and corresponds to a feature of the
environment. Each feature is characterized by a weight and a variance
parameter. The weight parameter represents the importance of a feature in the
localization problem, with respect to the other features. It is computed by
counting the cluster elements. The variance parameter, instead, represents the
precision of the estimation in the feature location referred to the fixed frame. It
is calculated as the maximum distance between the barycenter of the cluster and
the windows centers that belong to the cluster itself
- ’*/

*
1 3
I dl
I
i

1 1 I
1
t

..a,_ ..
.I
. ,.~,, j b . ~ .;-~.................
I .‘.j
i

Figure 2. Centers of the windows with three values of k parameter a d 1 0 0 b: 6 2 0 and C:P.

3. The implemented algorithm

The particle filter technique is well suited to deal with the GLA. It does not
make strong assumptions on the probability density of the robot pose after the
sensors readings. As a result, a particle filter can be used to represent complex
multimodal beliefs due to multiple possible poses.
The application of this algorithm requires an initialization phase, during
which a map of the environment is built by means of a features list. In order to
do that, the robot is placed on a set of well-known poses in a global reference
system and the clustering algorithm is applied to the corresponding scanner
readings, as described in section 2. The segmentation algorithm creates a list of
features that are points in the global reference system characterized by a weight
(number of scanning point linked to the feature) and a variance. These two
parameters are used to estimate the importance factor of the particles. The
particles filter algorithm is initialized by placing the samples randomly in the
map with the same initial heading of the robot. For each laser reading,
corresponding to the measure phase of the particle algorithm, the clustering
 569

process calculates the local map of the features. For each sample, the local map
is roto-translated in order to compare it to the global map, assuming that the
robot is located in the sample pose. It is then possible to compute the importance
factor of each particle, as described in the following section. This parameter
leads the re-sampling phase of the particle filter.

4. Estimation of the importance factor

The importance factor is the posterior probability of the robot to be in the
particle pose after the sensors readings. After the measure of the scanner laser,
the features of the local map are computed in the global reference frame for
each particle. In this way, the global list of features can be compared to the
features obtained by a scanner reading in the same reference frame. The first
step to obtain the importance factor is to compute the minimum distance of the
NJ features from the feature list (made of Nfl elements), for each particle k.

dkj= rninarg(Ilz, -z,II j = l...Nfl ,i = l...Nr (4)

Where ziis the i-th feature obtained from the scan and zj is the j-th feature
of the feature list. The importance weight of the k-th particle now can be
computed as follows:

Where Niis the number of scanning points that belong to the cluster i-th on
the feature list and oiis the variance of the i-th cluster.

5. The software architecture of the system

The first step of the software development was to give the possibility to
distribute the calculus in many computers, so a client-server architecture has
been chosen.
The other target to satisfy was the interoperability between the simulator and the
real environment. As a result the “Simulator Server” and the “3Morduc Server”
provides the same interface to the “Localization Client”.
Fig. 3 shows a block diagram of the whole system architecture [ 111.
The next subsection describes:
*Simulator
-3Morduc Server
570

*Localizationclier

Figure 3. The system arL...&ecture

6. Simulator
The environment simulator has been developed using GLscene components for
Delphi 7 that integrates Open Dynamic Engine [ 101.
The geometric dimensions of the simulated robot (wheelbase, radius of the
wheels, etc.) and the weight are the same of the 3Morduc platform.
The simulator models the measures coming from the scanner laser; in order to
obtain the intersections between rays and walls, a geometric algorithm has been
used. The scanner laser noise has been modeled using the classical beam model
of range finders, thus by adding hit, short, max and random noise [I].
Hit noise has a narrow Gaussian distribution characterized by a mean equal to
zero and a standard deviation ohit .
Short noise simulates the presence of unexpected objects and the probability of
range measurement in such situations and is described by an exponential
distribution.
Max noise simulates failures, i.e. when a ray is lost. This problem is modeled by
using a point-mass distribution centered at the maximum range of the laser.
Random noise, finally, models the unexplainable measurements as a uniform
distribution.
The simulator models also the encoders and examines the effect of two kinds of
noise sources. First a Gaussian noise, due to the encoder quantization, has been
7r
added to the measures, with mean p = 0 and standard deviation = -,
2Ne
where N , is the number of encoder pulse per turn. Moreover, it has been
considered the wheel-ground interaction modeled as a contact joint [ 111 (see fig.
4).
 571

Figure 4. Contact joint

This interaction prevents cylinder and plane inter-penetrating at the contact

point and forces the bodies to have an “outgoing” velocity in the normal
direction of the contact. Contact joints typically have a lifetime of one time step.
They are created when a collision is detected and then ~ e ~ a t edeleted.
l y
Contact joints can simulate friction at the contact by applying forces in the two
friction directions, perpendicular to the normal one.
The main parameters to characterize the contact are:
*mu: Coulomb friction coefficient ( mu E [&+a[); zero results in a frictionless
contact, and Go in a contact that never slips.
*mu2: Optional Coulomb friction coefficient for friction direction 2
(mu2 E [o,+a[ >.
*Bounce: Restitution parameter (Bounce E [0,1]); zero means that the surfaces
are not bouncy at all, 1 represents maximum bounciness.
*Bounce vel : The m ~ i m u mincoming velocity necessary for bounce (in d s ) .
Incoming velocities below this will effectively have a bounce parameter of zero.
*Soft erp and Soft cfm: normal “softness” of the contact. Fig. 5 shows a
screenshot of the simulator.

Figure 5. Simulator screenshot

7. The 3Morduc Server

The 3Morduc Server provides to the client the same command interface of the
Simulator application.
It implements the RS-232 communication protocol with scanner laser to
setting its parameters and obtains the measures.
The 3Morduc server also performs the communication protocol with the
ST10F269 motor control board to tuning the control parameter, set the speed of
572

the robot and get the encoders measures.

Finally, it adds a timestamp to the measures to permit, on the client, the
synchronization of the data.

8. The Localization Client

The Localization client requests the sensors measures to the server and allows
the user to choose between Simulator data or 3Morduc real measures.
Another important task of the client is the data synchronization using
timestamp, to have scanner laser measures consistent with the encoders data.
The localization client execute the GLA and allows the user to set the number
of the particles and the type of estimation of the importance factor.
The implemented weight estimation algorithms are:
Standard - uses the minimum distance of the laser data.
Mean Shift - uses the algorithm previously described.
The user has to load the map of the environment, an image that represents the
walls and the obstacles with black pixels on white background, before launching
the GLA.

9. Test and results

The first test in simulation mode has been prepared in a simple room of 1lm2.
Initially the clustering algorithm is used to obtain a list of features representing
the global map. The number of particles used in the GLA is 10000. They have
been initialized randomly in the map space as shown in Fig.6a with an initial
heading of -7c. The robot goes straight at constant speed and the scanner laser is
acquired at 4 Hz. The standard method, that uses only the minimum distance of
the laser data to estimate the importance factor, does not localize the robot after
23s (Fig.6a, Fig.6b and Fig.6~).The Mean shift method localizes the robot pose
in 5s by using only 20 scanner laser readings, the particles tends to condensate
in the correct position as shown in Fig.7.a, Fig.7.b and Fig.7.c. The particles are
concentrated in a very small region with a maximum diameter smaller than one
centimeter with a distance fiom the real position of 5mm.

Figure 6 . Particles distribution in classical GLA for three different time (a) t=O s, (b) t=5 s and
(c) t=23 s.
 573

s and (c) t=5 s.

The second test was performed using the real 3Morduc platform and has
been prepared in a room of about 24m2. The environment map has been
reconstructed by using six scanner data rows obtained when the robot was in
well-known pose. The number of samples in this case was 5000, uniformly
distributed in the room with initial heading of 4 2 as shown in Fig.8a. The
scanner laser data was acquired at 4 Hz sample frequency and the encoders
measures at 10 Hz. The GLA localizes the correct pose after 6 s, only 24
scanner laser data were needed to condensate the samples in a small zone with 7

Figure 8. Particles distribution in a real environment for three different time (a) t=O s, @) t=3 s and
(c) t=6 s.

10. Conclusion
This paper presents a GLA based on features extraction and particle filters,
applied to a aifferential drive robot. The environment map is a list of features
created using an adaptation of the MSA to the scanner laser readings. The
features are not geometric entities (corners, walls, objects, etc.) and most of all
they are very robust to noise. Moreover, two parameters are assigned to each
feature. The weight represents the relative importance of a feature with respect
to the other ones, while the variance is a measure of the feature precision in the
map. In this way, the map quality can be evaluated by considering the variance
of its features. The importance factor of the particles used for re-sampling is
computed by comparing the feature position on the local map to the features on
the global one. The local map is computed from the scanner laser measures
using the same segmentation algorithm: mean shift and clustering. From a
574

computational point of view this method is very lightweight and so it can be

used in real-time applications (150ms each step - Pentium 3GHz) . Finally, the
algorithm has a very fast convergence rate, as demonstrated from the reported
results.

Acknowledgments
This work has been partially supported by STMicroelectronics of Catania,
Automation and Robotics Team.

References
1. S. Thrun, W. Burgard, D. Fox, “Probabilistic robotics”, Ed. The MIT Press,
2005.
2. R.G. Simmons, D. Apfelbaum, W. Burgard, D. Fox, M. Moors, S. Thrun,
and H. Younes. 2000 , “Coordination for multi-robot exploration and
mapping”. Proc. of the National Conference on Artificial Intelligence
(AAAI), 2000.
3. W. Burgard, A.B. Cremers, D. Fox, D. Hahnel, G. Lakemeyer, D. Schultz,
W. Steiner and S. Thrun, “Experiences with an interactive museum tour
guide robot”, Artificial intelligence 1999, 114:3-55.
4. K. Konolidge and K. Chou, “Markov localization using correlation”, In
proceedings of the international Joint Conference on Artificial Intelligence
(IJCAI), 1999.
5. S. Thrun, “Bayesian landmark learning for mobile robot localization”,
Machine Learning 33,1998
6. F. Dellaert, D. Fox, W. Burgard and S. Thrun, “Montecarlo localization for
mobile robots”, Proceedings of the international Conference on Robotics
and Automation (ICRA), 1999.
7. D. Fox, W. Burgard, F. Dellaert and S. Thrun, “Montecarlo localization:
Efficient position estimation for mobile robots”, Proceedings of the
National Conference on artificial intelligence (AAAI), 1999
8. P. Jensfelt and H.I. Christensen, “Pose tracking using laser scanning and
minimalistic environment model”, IEEE Transaction on robotics and
Automation 17 pp.138-147, 2001.
9. D. Comaniciu and Peter Meer, “Mean shift analysis and applications”,
IEEE International conference Computer Vision (ICCV), 1999.
10., R.Smith, “Open Dynamic Engine ver. 0.039”, User Guide, 2004.
11. G. Muscato, D. Caltabiano, “A New Modular Architecture for the Mobile
Robot MORDUC: from the Hardware to the SLAM Algorithm”,
Proceedings of the 1st International conference on Dextrous Autonomous
Robots and Humanoids (DARH2005), 2005.
ASYNCHRONOUS LOCAL POSITIONING SYSTEM BASED ON
ULTRASONIC ACTIVE BEACONS AND FEED FORWARD
NEURAL NETWORKS*

PABLO ESTBVEZ, JUAN G. HERNANDEZ, JOSE CAPPELLETTO

AND JUAN C. GRIECO
Mechatronics Group, Electronics and Circuits Department, Simdn Bolivar University,
Sartenejas, Caracas 1020, Venezuela

This paper describes the design and test of an asynchronous 3D local positioning system
architecture for small indoor environments, using ultrasonic bursts emitters as active
beacons. Their times of arrival at the device are measured, defining its position in an
hyperbolic reference system, and these coordinates pass trough a feed forward neural
network which calculates the equivalent Cartesian coordinates. Various simulated setups
are used to evaluate the configurability of the system and one 2D setup is physically
implemented and tested, showing positioning error measurements for each case. The
proposed architecture shows itself easily configurable to a variety of setups, capable of
solving the positioning problem with small errors; its low computational cost together
with the asynchronous approach allows for a distributed implementation on small
devices.

1. Introduction
Spatially distributed applications, such as mobile robots and sensor networks,
make good use of localization systems, which allow them to know their current
position for tagging their sensor data and planning their navigation through the
working space. There are different options for it, depending on the size and
characteristics of the working space, the number of devices, and even the cost,
many of them referenced in [l]. The work presented on this paper is a part of an
investigation on underwater applications with multiple mobile robots, and the
special characteristics of that environment ask for a system not depending on
electromagnetic waves, since the attenuation factor is high in water; it should
also be distributed and asynchronous to be easily scalable to big number of
devices, and should be capable of 3 dimensional positioning. This paper
describes the design, implementation and tests of an LPS which meets most of
the requirements previously mentioned, even if not yet meant for underwater
environments. The proposed architecture relays on fixed ultrasonic emitters and

* This work was supported in part by Simon Bolivar University Research & Development Deanship.

575
576

one mobile receptor. By using the difference on the ultrasonic bursts time of
amval (DTOA), the position of the device is determined, with a feed forward
neural network performing the needed calculation.
The second chapter of this paper briefly describes the theory behind the
localization algorithm here used. Also, neural networks (NN) are presented as
an option for solving the equations system related to this kind of localization. In
the third chapter, its actual implementation is described. The fourth chapter
describes the set-up for the system tests, and shows the results gathered from
them, which are analyzed in the fifth chapter. The sixth one presents the
conclusions about the system performance and gives recommendations for
future works.

2. Theory of Operation

2.1. Hyperbolic localization

This technique is based on the difference in the travel time of signals (usually
radio or ultrasonic waves) between beacons and the device to be located.
Differences in the TOA of each signal depend on the difference in distance from
the device to each beacon. The time difference (and hence the distance
difference) between signals associated with to beacons defines a set of possible
locations described by an hyperbolic surface, creating an hyperbolic coordinate
for the device. In a system with m beacons Bb k=l ...m located at known
positions (xby&, each pair (Bi, Bi+,),i=l...m-1 generates an hyperbolic
coordinate hi.Notice that even if there are many other possible pairs, their time
differences are a combination of these ones, thus providing no new information
to the system. At Cartesian coordinates (x,y), the ithhyperbolic coordinate of the
device is:

In a three dimensional working space and as the number of beacons increases,

the solution space shrinks from surfaces to curves, then a set of points and,
finally, a single location. Also, constrains in the working space help on finding
the solution. For instance, a two dimensional space requires 4 beacons to locate
the device, and if only a section of it is observed, three beacons are enough, as
shown in Figure 1.
 577

Figure 1: Hyperbolic two-dimensions map for a three beacons system.

2.2. Neural Networks Solving Method

The problem faced here is finding a position given a set of measured time
intervals (or hyperbolic coordinates), which means finding a common solution
for a set of equations of the kind of (1). That equation has no explicit solution.
In hyperbolic navigation systems like LORAN and CHAYKA [2], maps are
used for graphical identification of the intersection. But for automated systems
numerical methods are needed, like the spherical interpolation presented in [3],
least-squares methods or the stochastic gradient method used in [4]for a setup
similar to this one. In this work, a Neural Networks approximation to the
problem is suggested. As can be seen in Figure 1, the problem is that of a
nonlinear continuous space transformation. The space described by the
“hyperbolic coordinates” of the device, given by the DTOA, has to be
transformed into a space described by Cartesian coordinates, and that is a
problem that feed forward neural networks can easily solve.
To train the neural network a set of input-output pairs should be provided.
These can be obtained from real measurements for a better result, but a
simulated approximation can also be used. Using (1) is easy to find the
hyperbolic coordinates for each Cartesian one, and these pairs (reversed) can be
used to train the NN.

3. Implementation

3.1. Ultrasonic Channel Sharing Protocol

For this system, distributivity is a requisite. Therefore, the mobile device
should do all of the calculations and needs to act as receiver of the signals
emitted by the beacons. Since all the emitters beacons send their ultrasonic
bursts at the same frequency, a protocol have to be established to share the
channel. In this case, the time division multiple access scheme proposed in [4]is
used.
578

3.2. Neural Networks algorithm

To make the system scalable, distributed and untethered, the feed forward
neural networks are required to run on s w l l microcontrollers with low
computational power. The neural network library used is designed using integer
numbers and optimizes the size of the variables and the kind of operations made
with them, to ensure small memory use and computing time [ 5 ] .
The estimate of the number of operations (NOO) per localization is computed
using relations (2) and (3). It is important to notice that all of these instructions
are performed over integer numbers, thus minimizing their computational
complexity. When using a standard neural network library, all these operations
become floating point ones and the number of multiplications is increased due
to the sigmoid function implementation.

Nod, = Number of add operations NaufpUfr

= Number of outputs

N,,,Ly = Number of mult. operations N,,,,, = Number of neurons

N,npu,s= Number of inputs in hiden layer

= NHfdenN ‘ (l + Ninpufs + Nourpurs )+ Nourputs (2)

Nrn~lfs = NHidenN ‘(l + N i ~ p u f s + Nourpurs )+ Noutputs (3) I5i
For the training process, a set of points is got from a grid in the working
space, and given the positions of the emitters, the receiving time intervals are
calculated for each one. A NN is created using the functions provided with the
library’s tool, and the data set is used to train it to solve the localization
problem. The resulting network is tested with Matlab@ and exported, to include
it on the microcontroller code. Also, real inputs-output pairs can be measured at
locations on the working space and used as part of the training data.

3.3. Ultrasonic Transceivers

For the physical implementation, SFR04 commercial ultrasonic range finders
were used as transceivers. These devices are designed to measure distances, but
in this case they are used only as burst emitters and device receivers. With the
receiver pointing upwards and the emitters fixed some distance above the
ground, a good enough signal could be received in most of the work area.
Nevertheless, these transceivers have a narrow lobe shape and even if they
could cover most of the area when pointing to its center, pointing to the
proximity of the device was needed in some areas, to avoid detecting reflections
instead of line of sight waves. This doesn’t affect the validity of the tests and
can be solved using omnidirectional emitters for final implementations.
 579

4. System tests and results

4.1. Simulation
The simulated tests ran in Matlab@. A data generator program was written to
create the vectors containing the Cartesian coordinates and the corresponding
DTOA for various arrangements of beacons. This program generated a training
data set from a regular grid and a validation data set of randomly chosen points.
Using Matlabo’s nntool and the [5] library transfer hnctions, different
networks architectures were tested, and their mean absolute error (MAE) to the
real position over the verification data set were measured’after 300 epochs. All
the NN have one hidden layer of either sigmoidal neurons (SN) or an integer
and computationally cheaper approximation from [5] (ISN), and a linear outptlt
layer. The Number of Operations (NOO) is calculated using (2) and (3) for the
ISN networks. The SN network uses Matlabo’s functions whose N O 0 are not
reported. The results of these tests are shown in Table 1.
I I xrnn I

4.2. Physical setup

In order to test the system in a real setup, the 31d and 4‘hconfigurations from
Table 1 were implemented. Both the emitter’s protocol and the receiver code
were programmed in separate microcontrollers. A 50 cm by 50 cm grid was
painted on the 2m by 3m work space, and the emitters were placed 1 meter
above the ground level to have better area coverage with the used transceivers.
580

This area was free of obstacles, but surrounded by a normal laboratory

environment, with walls and other ultrasound reflecting objects. The spatial
distribution of MAE in these tests can be seen on Figure 2.

12

10

+- 2.6m >-
Figure 2: Positioning errors (in cm) for two experimental setups.

Also, 55 measurements were taken on the work space and the FFNN of the 4"
test was trained for 10 additional epochs with this data to adjust it to the
conditions of the real environment. Then the test was repeated on points not
previously seen by the system, and the results are presented in Figure 3.

150 4

100 3
0
v
x
5n 2

0 1
50 100 I50 200 250
YIcnl
Figure 3: Positioning errors (in cm) for the equilateral setup trained with real measurements.

5. Results Analysis
The first test got very low values for the MAE, which could be expected for
such a big and complex FFNN. Nevertheless, given the computational economy
requirements of the system it becomes important to place our attention on the
 581

following experiences. Their use of the integer approximation of the sigmoidal

function yields to accuracy loss on the results, but requiring much less time to
compute them.
Tests 2-3 and tests 4-5-6 show an interesting property of the system which is
its graceful degradation with the decrease of the computational power. Even if
the accuracy of the systems gets worse when the FFNN is shrunk, it is still able
to get to a useful result.
Tests 3 and 4, and their corresponding physical implementations, show the
reconfigurability of the proposed system. With just a small change in the data
generator function and a fast training, the FFNN was able to work with a
different beacons-arrangement. Also, the physical tests demonstrate that the
system is able to generate appropriate position measurements in the real world,
even if trained only with simulated data. The MAE in these cases (7.lcm for test
3 and 8.3cm for test 4) were bigger than those obtained in the simulations,
possibly as a consequence of a too simplified training model. Also, small errors
on the positioning of the beacons affect the results, especially in their proximity
as can be seen on Figure 2. Nevertheless, Figure 3 (with a MAE of 2.2cm)
shows how these results can be greatly improved in the whole working space
whit just some measurements from the real setup and a few training epochs to
avoid over fitting, thanks to the inference capability of the NN. During the test it
was observed that the presence of objects or persons in the working space did
not affect the results, as long as the line of sight between the receiver and the
beacons was not broken, validating the timing scheme of [4]in those conditions.
The 5thto 7" tests show the capability of the system to work on a 3D space,
given enough beacons. The growth of the solution space yields to increased
positing errors for NN of the same size than those of tests 2 to 4.
Finally, the 8" to loth tests show the possibility of working on an
overdetermined scenario. The small improvement on the MAE from test 6 to 8
is not likely to be due to the extra information from the 4" input, since in the
simulation it is just a combination of the other three, but due to the extra
network parameters. Nevertheless, it shows that it is feasible and easy to include
beacons that over determinate the system, which might be useful in noisy
environments. This is proved in test 9 and 10, where the presence of the 5Ih
beacon improves the accuracy of the system when fed with noisy input data.

6. Conclusions and future work

The simulated tests proved the capability of the proposed architecture to solve
the positioning problem in a variety of conditions, resulting in an adequate
translation of the DTOA to the Cartesian coordinates. The algorithm showed
itself useful and effective to do 2D and 3D local positioning, with errors whose
582

magnitude can be handled with the size and complexity of the FFNN used.
Nevertheless, when doing so the computing requirement of the system should be
taken into account.
Its capacity to work in presence of reflecting objects makes of it a real
possibility for indoor environments as long as line of sight between the beacons
and the device can be assured across the work space. The system is easily
configurable to a specific setup with just a computer simulation, and a small
number of measurements improves remarkably its performance.
Finally, the asynchronous approach and the low computational requirements
of the solution allows for its application to environments with many robots
sharing a working space, since each one of them can do its own position
estimation at no big computational, power or even economic cost.
In order to improve the accuracy of the system trained with simulated data, it
is suggested to develop a better model of it, taking into account all of the
variables that affect the velocity of sound like temperature and air pressure,
which can be measured by the device while working and included in its
calculations. The optimal placement of the beacons should also be studied since
their positions affected the error distribution and values. For a real and useful
implementation of the system, omni-directional transceivers have to be
implemented. The use of a CDMA channel sharing would greatly improve the
position computing time and robustness of the system, specially for
overdetermined scenarios.
Given the framework of this investigation, it’s planned to test the system in
the future in underwater environments using the appropriate transducers.

References
1. J. Hightower and G. Boniello, “Location systems for ubiquitous
computing,” IEEE Computer, vol. 34, no. 8, pp. 57-66, August 2001.
2. A. R. Jimenez, F. Seco, R. Ceres, and L. Calderon, “Absolute Localization
using Active Beacons: A survey and IAI-CSIC contributions,” Institute for
Industrial Automation, CSIC Madrid.
3. J. Smith and J. Abel, “The spherical interpolation method of source
localization,” IEEE Journal of Oceanic Engineering, vol. 12, January 1987,
pp. 246-252.
4. Fermin, Leonardo; Medina, Wilfredis; Chinea, Ana; Veloz, Nicolas;
Grieco, J.. “A Low Cost Local Positioning System using Ultrasonic
Sensors”. 9th International Conference on Climbing and Walking Robots
(CLAWAR 2006). Brussels, Belgium. September 2006. CD.
5. P. EstCvez, “Implementacih y prueba de algoritmos para reconocimiento
de context0 sobre plataformas smart-its’’ BSc. thesis, Dept. Electron. Eng.,
Sim6n Bolivar University., Caracas, Venezuela, 2005.
 CONTACT PROCESSING IN THE SIMULATION OF CLAWAR

JUHASZ', KONYEV~,RUSI", SCHMUCKER~

'Department Virtual Engineering, Fraunhofer Institutefor Factory Operation and
Automation, Sandtorstrasse 22, 39106 Magdeburg, Germany
'Institute for Electrical Energy Systems, University of Magdeburg, P.B. 4120, 391 06
Magdeburg, Germany

Contact processing, including collision detection and collision response, is one of the
most difficult, but most important areas in simulation of the multi-body systems.
However, the most widespread multi-body simulators, like Matlab/SimMechanics or
Modelica/Dymola, don't support the contact processing. Other multi-body simulators,
like Vortex or ODE, support the contact processing, but are more limited in the rest of the
functionality. This paper presents the implementation of the contact processing in both
Matlab/SimMechanics and ModelicdDymola and the comparison of the implemented
functionality with Vortex. It was performed through an example of contact tasks for a six-
legged robot.

1. Introduction
In mechanical systems certain machine elements usually interact with each
other. When a mathematical model of such system is designed, the interactions
between the parts can be divided into two following categories: -Mechanical
joints are used for defining permanent constraints of motion. -Mechanical
contacts are almost instantaneous, typically short-time interactions caused by
non-penetration contact forces arising between the bodies in the model. The
forces occur when the surfaces of bodies touch each other. Two major
phenomena occur in the mechanical contacts: collision contacts (causing
collision response forces) and friction contacts (causing static or dynamic
friction forces).
Contact processing is a difficult task [1,5,6]. The bodies can move in a
complicated way, and they can have complex geometries. In the case of
contacting bodies, the penetration of them must be prevented. There is a tradeoff
between efficiency and accuracy. Accurate methods for computing contact
forces are based on finite element methods. Such methods are based on
subdivision of bodies into very small fragments. The surfaces of two colliding
bodies are to be covered by a mesh and the relevant forces in the contact are to
be computed for each point on the mesh. The resulting forces can be defined by

583
584

integration of all forces acting on the contact surface. These methods are
implemented in software packages for FEM-analysis (ANSYS, Nastran etc) or
in multi-body simulation (MSC Adams etc). Experiments [2] show that these
methods are accurate, but require tremendous computing resources and
therefore are very slow. However, many simulation applications do not require
extreme accuracy and additional assumptions are taken into account providing
high simulation speed, but decreasing the accuracy. As a matter of fact, different
assumptions lead to different computation methods but with the same (or nearly
the same) computation results. In such cases, it’s not important for the
application what assumptions and methods were used.
The goal of this paper is to present the ways of adding contact processing
(chapter 2) to existing mechanical multi-body simulator Matlab/SimMechanics
and to compare them with contact processing in Vortex, which is characterized
by internal optimization loop with considering of energy and impulse
conservation law. The results are presented though the example of contact tasks
for a six-legged robot (chapter 3).

2. Contact Processing
Implementation of the contact processing is based on the following steps.
(1) Mechanical models, which describe physical bodies, should be extended to
describe contacting physical bodies. (2) There should be a routine that can
detect collisions and can return detailed information regarding contact
parameters, such as contact points and their velocities. (3) A special routine
should calculate the contact response from contact parameters. (4) Each of these
components should have an interface that allows replacing its implementation
without doing major redesign of the other components.
Figure 1 presents four basic components of contact processing, which are
D
ve
o
r
df
m h
t
ed
sc
i
u s
o
s
in
ab
ov
ea
n
bd
e
w
l
dl
ei
sb
ci
e
rdmo
e
r
de
e
a
a
ld
e
t
l
t
i
.
r

. Update of
System Dynamica:
new positions and
Orientationsof the bodies
. Contact Response:
bady penetration
prevention.contact force
calwlation or body
impulse change

integration Loop: OptlmizaUon Loop:

of impulse and energy

Collision c O n t s d area I “0l“rna I

Detection: direction.
”0 I yes penetrationdepth I

Figure 1. General scheme of contact processing in the mechanical multi-body simulation

 585

2.1. Collision Detection

The collision detection between bodies, which are indirectly described, for
example using sets of points in the space, means that the contact points or
trajectories between two bodies are to be defined.
In the first part (collision culling), the pairs of non-colliding objects should
be excluded. These can be performed using the methods of space division like
Quadtree/Octree, BSP-Tree, Sweep-and-Prune. The whole space should be
divided and the potentially non-colliding objects should be excluded.
In the second part (broad collision detection), the possibility of two objects
collision by means of so-called Bounding Volumes like AABB, sphere, OBB, k-
DOP should be defined. These covers simplify the complex geometry objects
and make the collision detection simpler and faster.
The third part (exact collision detection) defines the collision between the
contacting objects. Spatial, herarchical data structure, as AABB-tree, OBB-tree,
sphere-tree or k-DOP-tree should be used for faster collision definition.

2.2. Contact Parameters

The intersection test follows the third part of collision detection and finds
the geometrical contact parameters, as contact plane, contact volume, contact
normal and penetration, with the aim of the following methods: intersection
point of ray and sphere, intersection point of ray and plane, intersection point of
ray and triangle, intersection line of two planes, intersection line (joint) of two
triangles. The presented on the market algorithms (Lin-Canny Closest Features
Algorithm, I-/Q-COLLIDE, V-Clip, OBB-Tree, QuickCD, KDS, GJK, GJK-
based EPA) combine “collision detection” and “contact parameters detection”
tasks [7-111 and are implemented in the leading softwares, e.g. SWIFT, SOLID,
ODE and others.

2.3. Contact Response and Update of System Dynamics

The contact response seems to be the most problematic and controversial
part of the contact processing, since many computation approaches exist, which
require different input information and may produce quite different numerical
results. The following two methods are commonly used in the contact
processing: the impulse-based method and the force-based method. Both assume
that the bodies are rigid. The update of system dynamics is closely connected
with calculation of contact response and therefore both parts should be
explained together.
586

2.3.1. Impulse-based Approach

The impulse-based approach uses collision impulses between the bodies
and changes the velocity vector of the bodies during the contact [3,4]. This
method based on an impact law such as Poison’s hypothesis. It considers the
impulse conservation law and operates with the impulses of the colliding bodies
before and after the collision as well as with the restitution coefficient of
materials.
The main advantages of t h s method are that only a few constants are
needed for description of the impact law and that the integrator step size is not
influenced by the response calculation because it is performed during an
infinitely small time instant. However, since the velocity is not continuous in the
impulse-based model, the traditional ODE solvers can’t be used. The continuous
integration process in the solver should be stopped at the instant of collision and
should be resumed with a new velocity. The impulse-based approach can be
easily used in MBS-based models if the collision impact on the other bodies in
the system is negligible (i.e. in the system of free-flying bodies). In the other
words, this approach can’t be used in the cases of a static objects and structures
consisting of several bodies connected by joints. Furthermore such idealized
impact laws are only useful for stiff collisions. These properties restrict the
applicability of the impulse-based method of the dynamical analysis.

2.3.2. Force-based Approach

An alternative approach of contact processing in multi-body mechanical
systems is based on the force and torque model of collision. It is assumed that
the contacting bodies penetrate each other and the separation forces are caused
by this penetration. These forces try to prevent further penetration and to
separate the contacting bodies.
The calculation of contact force magnitude is difficult task and is
sometimes not motivated by physics, but rather by numeric analysis. The overall
result of collision should match physical laws (i.e. preservation of impulse, and
preservation of energy). In addition it should be chosen so smooth that
numerical methods used in simulation could handle these functions. The many
existing methods for the calculation of the force in the mechanical joints and
contacts are divided into two following groups:
Force based methods with Lagrange multipliers formulation models the
mechanical constraints (contacts and joints) with the reactive forces, which
are presented as Lagrange multipliers 1. The constraint forces perform no
work on the environment and the physical meaning of the mechanical
 587

contact is lost. The mechanical interaction of the bodies caused by the

contact is represented by these reactive forces I , which should be optimized
between the simulation steps in the additional optimization loop (s. Fig.1)
under consideration of the energy or/and impulse preservation laws.
Force based methods with penalty formulation models the mechanical
contact with the strong poss. nonlinear spring. The active contactlfriction
forces (s. Fig.2) perform work on the environment and the physical
background of the mechanical contact is not lost. The mechanical
interaction of the bodies caused by the contact is represented by the active
forces FcONTACT and FFRICTION, without any additional optimization between
the simulation steps.

Figure 2. Force based methods with penalty formulation for contact and friction forces calculation.

There are many different equations for determining the contactJfriction

force magnitudes, which depend on the penetration depth p , on the penetration
velocity dp/dt, on the fnctional penetration I, and on the Coulomb friction
coefficient p. For the purpose of this work the following equation system is
chosen for the calculation of the force magnitudes, which is very stable in the
wide range of the simulation sample time and corresponds to all above
mentioned requirements. Eq. I describes the contact force magmtude depending
on the stiffness ScoNTAcT in the contact area, the restitution factor E of the
materials and the collision velocity v c o L L I ~ I ~
atNthe first time instant of
intersection:
588

The main advantages of the force-based approach are the simplicity and
the possibility of using it for stiff and soft contacts. This approach works
reasonably well if several contact points are also present at the same instant
time. The disadvantage of this approach is that the integrator step size should be
reduced in the contact phase in order to catch the rapidly changing contact
forces and torques. And similar to the impulse-based method there is necessary
to choose the contact parameters (spring, damping, restitution) because the
contact force is proportional not only to the penetration depthkelocity but also
to the contact area and the contact volume.
Update of system dynamics take place each integration step depending of
the acting contact forcedtorques.

3. Simulation of Contact Tasks for Six-Legged Robot

The force based contact processing method with penalty formulation,
represented in chapter 2, has been implemented according to Fig.1 into the
multi-domain simulation environments Matlab/Simulink (s. Fig.3) by means of
Solid as collision detection software and has been compared with the force
based contact processing method with Lagrange multipliers formulation,
implemented into multi-body simulation environment Vortex. The six-legged
robot “SLAIIU”, developed at the University of Magdeburg and the Fraunhofer
Institute for Factory Operation and Automation, has been used as a test subject.
Two modes, staying on the ground and moving on the surface, have been
investigated and compared in the mentioned simulation environments (s. Fig.4).

Figure 3. Modular six-legged robot “Slair2” with articulated body in the multidomin simulation
environment “Matlab/Simulin~.
 589

Investigations show that the contact processing in Vortex is very fast,

robust and reality-near. It is caused by the first-order integration method with
adaptive step and by the optimization loop. However in dynamic mode the
calculation of the contact forces is not precise, because the number of the
optimization steps is limited to 35 by developers.
The developed contact processing environments in Matlab/Simulink are
free from aforementioned disadvantages, can use the integration methods of
higher order, and are sufficiently precise by calculation of the contact forces,
because Eq.1 has been laid out under consideration of the energyhmpulse
preservation law.

Figure 4. Comparison of the normal forces under the feet of the six-legged robot “Slair2” staying on
the ground and then doing four steps on the surface in (top) Vortex and (down) SimMechanics/Solid.
590

References
1. V. Engelson. Integration of Collision Detection with the Multibody System
Library in Modelica. Thesis, PELAB, IDA, Linkoping University, 2000.
2. D.. Fritzson, P. Fritzson, P. Nordling, T. Persson. Rolling Bearing
Simulation on MIMD Computers. Int. Journal of Supercomp. Appl. and
High Performance Computing, 11(4), 1997.
3. B. Mirtich: Impulse-based Dynamic Simulation of Rigid Body Systems.
Ph.D. thesis, University of California, Berkeley, 1996.
4. P. Zhang. Physically Realistic Simulation of Rigid Bodies. Thesis,
Department of Computer Science, Tulane University, 1996. Available via
http:llwww.eecs.tulane.eddwwwlZhan~/.
5 . M. Otter, H. Elmqvist, J.Diaz Lopez. Collision Handling for the Modelica
MultiBody Library. 4th International Modelica Conference, pp.45-53,
March 2004.
6. N.Galoppo, M.Otaduy, P.Mecklenburg, M.Gross, M.Lin. Fast Simulation
of Deformable Models in Contact Using Dynamic Deformation Testures.
ACM SIGGRAPH Symposium on Computer Animation, 2006.
7. M. C. Lin: Efficient Collision Detection for Animation and Robotics. Ph.D.
thesis, University of California, Berkeley, 1992.
8. K. Chung and W. Wang: Quick Collision Detection of Polytopes in Virtual
Environments, , ACM Symposium on Virtual Reality Software and
Technology 1996, 1-4, July, 96, University of Hong Kong, Hong Kong.
9. D. Schmalstieg, R. F. Tobler: Real-time Bounding Box Area Computation.
Institute of Computer Graphics, Vienna University of Technology, 1999.
10 J. Erickson, L.J. Guibas, J. Stolfi and Li Zhang: Separation-sensitive
collision detection for convex objects; Proceedings of the tenth annual
ACM-SIAM symposium on Discrete algorithms, Pages 327 - 336, 1999.
11 G. van den Bergen: Collision Detection in Interaction 3D Environments.
2003.
 CREATING A GESTURE RECOGNITION SYSTEM BASED ON
SHIRT SHAPES

PAVEL STAROVEROV, SILVIA MARCOS, DMITRY KAYNOV, MARIO

ARBULU, LUIS CABAS, CARLOS BALAGUER
Robotics Lab, Department of System Engineering and Automation, University Carlos III
of Madrid, Av. Universidad 30, Legangs 2891 1(MADRID), Spain

The present article describes the gesture recognition system that has been developed at
the University Carlos I11 of Madrid. The system’s functionality is based on the
assumption that the person that wants to interact with the robot should wear a shirt which
characteristics must have been previously passed on to the program. The algorithm used
for the recognition is presented in detail. The results of a number of tests are presented
and discussed along with the suggestions for hture improvements.

1. Introduction

Gesture recognition can be a very helphl tool that can make a welcome addtion
to the interaction abilities of any service robot. Humanoid robots would benefit
from gesture recognition even more, because t h s natural way of communication
that is often used between humans, can open the door to a new level of
interaction, that can be logically continued by imitation and learning.
The objective of this work is to propose a gesture recognition system for the
Rh-1 humanoid robot, a system that is easy to implement yet presents an
acceptable performance. Rh-1 is a humanoid robot developed at the University
Carlos I11 of Madrid (Fig. 1).
It has 2 1 degrees of freedom (and 2 more degrees of freedom of the built-in
camera). It measures approximately 130 cm and weighs in about 45 kg. It
already can speak, recognize voice commands, faces and “T-shirts”. Those
abilities were discussed in one of our previous works [lo]. Somewhere in the
present proceedings you may find another article concerning the recently added
sound source detection system.
The present article is organized as follows: in Sec.2 some information
concerning related works can be found, Sec.3 describes the way the system
works, experimental results are presented and discussed in Sec.4. Section 5
concludes the paper.

591
592

R
F
g
1
.
.
i
h
o1
r
-b
o
ap
n
t
d
se
v
mr
e
ar
lem
hb
t
e
R
e
o
sr
fo
c
bL
s
o
i
at
b

2. Related works
Recently there has been a lot of interest towards the field of gesture recognition.
One the works that we would like to mention is by Kojo et aE. [I]. It deals with
an advanced gesture recognition system for the HRP-2W robot, which is the
wheel version of the famous HRP-2P humanoid. The algorithm used is based on
the suggested Pro~o-symbo~ Space and also uses Hidden Markov Models
( H M ) , which is a widely used method for gesture recognition [2]-[5]. Calinon
and Billard in their research go further and use HNIMs for progra~mingby
demon~tration[ 6 ] .In their work [7] they also use an algorithm combining PCA,
ICA and HMM. Later Calinon and Billard and some colleagues inspired by
them made another step forward and used a similar method to reproduce the
movements on the robot [S, 91.
The algorithms used nowadays for gesture recognition are quite advanced,
but the downside is that they are not very easy to implement. We wanted to use
an algorithm that would be easier to implement and also make use of the
software created earlier.
 593

3. Development of the system

The design of a robust gesture recognition system can present significant
challenges to the researcher. Mostly this happens because of the fact that many
gestures are shown in 3D space, thus creating the need to extract spatial
information from 2D images. Our robot doesn’t have a stereo vision, it has only
one camera, so working with 3D gestures would be very hard. So it was decided
to limit the gestures that can be recognized by the robot. First of all, only the
upper body gestures will be considered. Those gestures are often used in the
signals of the traffic police, referees at sporting events and flight directors on the
airport runways [ll]. The lexicon of those signals is always finite, so the
number of upper body gestures to recognize can be limited by several poses. It
was decided to select the following gestures (Fig.2): 1) no gesture (both hands
down), 2) left hand stretched out to the left, 3) right hand stretched out to the
right, 4) left hand up, 5 ) right hand up, 6 ) both hands stretched out to the sides,
7) both hands up. At the moment the system works only with static gestures.

Fig.2. Processed sample pictures with the gestures that can be recognized by the program.
594

Another assumption that was to be made is that the person that wants to
interact with the robot should wear a shirt of a color that must be previously
passed on to the robot. It seems quite normal because road police, referees and
flight directors usually wear some kind of uniform to be distinguishable from
other people. Some initial restrictions do apply to secure the system’s acceptable
performance, namely, the shirt must be of a color that is different from the
objects in the background and the illumination during the calibration and the
recognition phases must be as close as possible.
It was decided to modify the “T-shirt based” image recognition system
described in one of our previous works [lo] to detect gestures using a similar
straightfo~ardalgorithm. Our idea was to recognize gestures using geometrical
properties of shirts (various types of clothes can be used but all of them will be
referred to as shirts for simplicity). Shortly, that system is capable of finding T-
shirts or similar pieces of clothes in the images using color information and it
can also calculate the distance to the person. The execution of the program can
be controlled by voice commands, and the result of the processing can be
comm~icatedto the user by the robot in the spoken form, too.
The gesture recognition process is divided into two stages: image
processing and gesture classification. Image processing tries to separate the
object of a known color from the background. The algorithm used is nearly the
same that has been used in our previous effort, although some improvements
have been made in order to provide a better object separation. In Fig.3 you may
see the result of processing of two sample images (on the left>,the first image is
with a red shirt and the second image is with a green one. On the right you may
see the resulting images that can be recognized correctly in the second stage.

Fig.3. Two sample images and the result of their processing.

 595

In the second example it can be seen that the right hand was clipped
because the person was too close and was not at the center of the image. AS we
will see later, in this case there is a high probability of a recognition error.
After the image has been extracted, the geometrical parameters of the
extracted object are analyzed. First of all, we look at the heighdwidth ratio. If it
is between 0.95 and 1.2 then it is supposed that there is no gesture (case 1). If it
is less than 0.65 then it is supposed that both hands are stretched out to the sides
(case 6). If it is between 1.2 and 3 then we verify whether both the top right
pixel and the top left pixel are at the same side of the image. If it is true then an
extra check is made to make sure that the mirrored hand area is black, after that
the corresponding gesture is chosen (cases 4 or 5, left or right hand up). In the
contrary case an extra check is made to make sure that the area between the
hands is black and after that it is supposed that both hands are up (case 7). If the
heightlwidth ratio is between 0.65 and 0.95 then we verify which pixel, the right
top one or the left top one, is farther from the vertical centroid and the
corresponding gesture is chosen (cases 2 or 3, one of the hands stretched to a
side). If it’s more than 3 then it’s considered a recognition error. The values of
the heighdwidth ratio may be adjusted for various persons.
Having detected a gesture, the robot may react to it in a way that we would
want to, for example, perform some action like stopping, making one step,
turning left, right, around, etc. It is the kind of response action one would
expect, for example, when following the signals of police directing traffic. Also,
the robot’s number of degrees of freedom lets it imitate all of the given gestures
if needed, although the robot’s response movement has not been implemented
yet in any form. Currently the robot just tells the person that a certain type of
gesture has been detected.
The suggested algorithm has a lot of assumptions and seems to be too
simple to be robust, but let us make the final decision after seeing the
experimental results.

4. Results and discussion

In order to evaluate the performance of the system, two series of tests have been
made. Two persons were repeating all the gestures three times, one of the
persons was wearing a red shirt, and the other was wearing a green one. All the
tests were made under normal lighting conditions. The persons intended to
remain parallel to the robot and to be reasonably far from it to enter in the image
completely. However, due to the dimensions of the room where the test was
conducted, it was not always possible. The results are presented in Table 1.
596

Table 1 . Test results for two persons repeating all gestures three times
(+ means a positive result, - means a negative one).
GESTURE TYPE RED SHIRT RESULTS GREEN SHIRT RESULTS
No gesture +++ +++
Lett hand stretched +++ tt-
Right hand stretched +++ ++-
Left hand up +t+ i++

Right hand up t+t ++-

Both hands stretched U- +++
Both hands up +tk ++-
It can be seen that the results of the red shirt are a bit better, resulting only
in one miss when both hands were stretched out but the program recognized
only one hand stretched. It was due to the fact that the room where the tests
were performed was rather small, and the person’s hands did not enter in the
image completely. In the case of the green shirt, there was a similar case when
the right hand was stretched out. There have also been two misses because of
the person’s inclined pose during the test (it is important to stand up straight)
and one m i s s due to the changed lighting conditions. In general, there have been
3 1 hits and 5 misses out of 42, which can be considered a satisfactory result. It
can be said that the system has an acceptable performance for our type of
application. It also has several advantages, namely, it is easy to implement and it
doesn’t need any training.
Nevertheless, the described system makes a lot of assumptions, depends
greatly on the calibration quality, on the lightning conditions and other factors.
Also, currently it works with static images only. All of this can be a ground for
improvements. First of all, it is recommended to use an algorithm based on
HMM.Although it is more difficult to implement, it may also provide better
results. Another important improvement would be the possibility to process
online video instead of static images. And finally, it is important to program
robot’s movements in response to those gestures, be they imitation movements
or some other movements. Later it would be interesting to make some kind of
learning by demonstration as suggested in [6] and [7].

5. Conclusion
In the present paper a gesture recognition system developed at the University
Carlos I11 of Madrid was described. A simple geometrical method is used to
recognize shrt shapes. The algorithm was described in detail, the experimental
results were presented and discussed. The performance of the system was
considered acceptable. Finally, several improvemenis have been suggested.
 597

Acknowledgments
We would like to thank the members of the Robotics Lab of University Carlos
I11 of Madrid for their cooperation and suggestions. The present work was
supported by CICYT (Comisi6n Interministerial de Ciencia y Tecnologia).

References
1. N. Kojo, T. Inamura, K. Okada, M. Inaba “Gesture Recognition for
Humanoids using Proto-symbol Space”, 2006 6th IEEE-RAS International
Conference on Humanoid Robots (Humanoids 2006), pp.76--8 1,2006.
2. S. Eickeler, A. Kosmala, G. Rigoll, “Hidden Markov Model Based
Continuous Online Gesture Recognition”, In Proc. Int. Conference on
Pattern Recognition (ICPR), Brisbane, 1998, pp.1755-1757.
3. Rigoll, G., Kosmala, A., and Eickeler, S. “High Performance Real-Time
Gesture Recognition Using Hidden Markov Models.” In Gesture and Sign-
Language in Human-Computer Interaction, pages 6980, 1998
4. R. Yang, S . Sarkar, “Gesture Recognition using Hidden Markov Models
from Fragmented Observations,” CVPR, pp. 766-773, 2006 IEEE
Computer Society Conference on Computer Vision and Pattern Recognition
- Volume 1 (CVPR’06), 2006
5. J. Yang and Y. Xu, “Hidden Markov Model for Gesture Recognition”, tech.
report CMU-RI-TR-94-10, Robotics Institute, Carnegie Mellon University,
May, 1994
6. Calinon, S. and Billard, A. “Gesture Recognition and Reproduction for a
Humanoid Robot using Hidden Markov Models.” AMVPASCALAM2A44
Workshop on Multimodal Interaction and Related Machine Learning
Algorithms. 2004
7. Calinon, S. and Billard, A. “Recognition and Reproduction of Gestures
using a Probabilistic Framework combining PCA, ICA and HMM. “In
Proceedings of the International Conference on Machine Learning (ICML).
2005
8. Calinon, S. and Billard, A. “Incremental Learning of Gestures by Imitation
in a Humanoid Robot.” in Proceedings of the ACM/IEEE International
Conference on Human-Robot Interaction (HRI) 2007.
9. T. Asfour, F. Gyarfas, P. Azad and R. Dillmann “Imitation Learning of
Dual-Arm Manipulation Tasks in Humanoid Robots”, IEEE-RAS
International Conference on Humanoid Robots (Humanoids 2006), Genoa,
Italy, December 2006
10. P.Staroverov; M.Arbulu; L.M.Cabas; DKaynov; C.Perez; C.Balaguer. “A
Voice Controlled Image Recognition System.” Proc. of the 9th International
Conference on Climbing and Walking Robots. Brussels. Belgium. Sep,
2006.
598

11. Sclaroff, S., Betke, M., Kollios, G., Alon, J., Athitsos, V., Li, R., Magee,
J., and Tian, T. 2005. Tracking, Analysis, and Recognition of Human
Gestures in Video. In Proceedings of the Eighth international Conference
on Document Analysis and Recognition (August 31 - September 01, 2005).
ICDAR. IEEE Computer Society, Washington, DC, 806-8 10
 DESIGN AND DEVELOPMENT OF MICRO-GRIPPING
DEVICES FOR MANIPULATION OF MICRO-PARTS

Z.W. ZHONG, S.K. NAH, S.H. TAN

School of Mechanical and Aerospace Engineering, Nanyang Technological University,
50 Nanyang Avenue, Singapore 639798, Republic ofsingapore

In this study, design, fabrication and tests of micro-grippers were carried out. The
geometry design and the material stresses were considered by means of the finite element
analysis. The simulation model was used to study the profiles of von Mises stresses and
deformation. Micro-gripper prototypes were made using various machining methods.
Micromanipulation tests were conducted to evaluate the performance of the micro-
grippers and confirm potential applications of the micro-grippers in handling micro-
objects. The simulation and experimental results have proven the good performance of
the micro-grippers.

1. Introduction
An effective mechanical micromanipulator should have the ability to grasp
objects of different shapes with high positioning accuracy. The manipulators
should be able to accurately control grasping forces in order to avoid any
damage to the small-size delicate objects, which are less than 1 mm in size [ 11.
Micro-scale technologies have been developed for many applications in
electronics, information technology, optics, medicine and biology covering
areas such as diagnostics, drug delivery, tissue engineering and minimally
invasive surgery [2]. A microgripper is one key element in micro-robotics and
micro-assembly technologies for manipulating micro-objects. Various
prototypes of microgrippers have been developed using electrothermal actuators
[6-91, electrostatic actuators [ 10-1 51, piezoelectric actuators [ 16,171,
electromagnetic actuators [ 181 and shape memory alloy actuators [ 19-22].
Micro-parts with major dimensions less than 0.1 mm are often fragile and
can be easily damaged during gripping, and thus special grasping techniques are
required. The specifications to realize a microgripper are quasi-static motion to
have high accuracy in micro-positioning, a large-stroke to grasp the maximum
types of object, and the use of special actuation method [23].
In this study, design, fabrication and tests of micro-grippers were carried
out. The geometry design and the material stresses were considered by the finite

599
600

element analysis. The simulation model was used to study the profiles of von
Mises stresses and deformation. Micro-gripper prototypes were made using
various machining methods. Micromanipulation tests were conducted to
evaluate the behavior of the micro-grippers.

2. Basic Studies

2.1. Actuation of Mechanism

A simple-structured microgripper with piezoelectric actuation was designed.
The actuator utilizes the shear deformation of piezoelectric elements to drive the
compliant mechanism of the microgripper. The entire microgripper consists of a
compliant gripper mechanism, a shear-mode piezoelectric actuator and a support
base. An exploded assembly view of the microgripper system is shown in Figure
1.

Piezoelectric

Figure 1. Assembly view of the microgripper system.

The piezoelectric actuator drives the movable arm of the compliant gripper
mechanism back and fro, thus providing the desired actuation to grasp and
release the target micro-parts or biological cells. For the assembly process, the
piezoelectric actuator is secured to a solid base surface. Epoxy glue provides a
strong, flexible and non-conducting bond. The flexible nature of a glue bond
also eliminates the possibility of vibration-induced fatigue during operation.
 601

2.2. Finite EIerttent AnaIysis of Microgripper

FEA software ANSYS was used as a modeling and calculation tool. The
simulation work involved the use of structural elements to model the compliant
mechanism and also the use of contact elements to model the interacting bodies.
Visual animations and numerical calculation results were obtained from the
program. The basic objective of the proposed compliant mechanism design is to
obtain desirable output displacements of the tip of the microgripper with known
input actuating displacements. The compliant mechanism is an elastic
mechanism that performs its kinetic function through its intended flexibility,
which is enabled by the notch hinges on the parallel arm links. Thus the
mechanism performance largely depends on its material properties and the
geometrical design. The design is a planar, monolithic mechanism consisting of
parallel links, flexural notch hinges and actuating links.
The maximum stress and the bulk of the stresses are concentrated at the
flexural hinges of the compliant mechanism. This is expected as the flexural
hinges experience the most bending deformations within the whole mechanism.
When the FEA model has a parallel arm length of 1.5 mm and the input
displacement is 10 pm, the value of the maximum stress obtained at the flexural
hinges is 828 MPa, which is within the elastic region of the mechanism material.
For the simulation study, the variables are the length of the parallel arms, I,
and the input displacement value. The ratio of the output displacement and the
input displacement is the geometrical advantage. The graph in Figure 2
illustrates that the longer the parallel arms, the greater the geometrical advantage
is. The geometrical advantage provided by the microgripper design is in the
range of 3.5 to 5.5 depending on the length of the parallel arms.

2 -3. Contact Simdation of Microgripper and Micro-Part

The circular object is flexible while the microgripper tips are rigid. Thus upon
contact, the circular object is deformed while the microgripper tips still retain
their shape. For one simulation, material 1, which is the material of the
microgripper, is defined as Foturan glass. Material 2, which is the material of
the target object, is defined as polystyrene. To avoid slipping, a coefficient of
friction value of 0.2 is included in the simulation. Two additional contact
element types are also defined to study the contact relationship.
When a deformation of 1 pm is applied on both sides of the circular
polystyrene object, the contact pressure is obtained as shown in Figure 3. The
maximum pressure occurs at the center of the contact region and decreases
exponentially to zero towards the outer end of the contact region. The
602

corresponding stress experienced by the object can be expresses by a parabolic

curve with its peak value of 436 MPa at the centre of the contact area, which is
the location where the target object experiences the maximum stress. The stress
on the target object drops rapidly to zero with increasing distance away from the
centre of the contact area.

3. A Prototyped Gripper Actuated Using a PZT Actuator

3.1. Design and Fabrication

The basic study results of Section 2 have led to several microgrippers fabricated
and tested. The dimensions of one of the microgrippers fabricated were
designed in such a way that micro-manufacturing technologies could be used,
leading to low cost and an easy adaptation to different micromanipulation needs.
The microgripper was fabricated from a single piece of spring steel using a
micro-wire electrical discharge machine (EDM). Another material used to
fabricate was aluminium. These materials were selected because they are
common engineering materials and are easy to be machined by EDM.
The micro-wire EDM technology can generate the microgripper with good
precision. With achievable dimensional tolerances o f f 5 pm and the capability
to use 420-pm wires, the wire cutting method is suitable for micro machining.
The microgripper prototype is 36 mm long, 30 mm wide and 3 mm thick.
The prototype of the gripping device is shown in Figure 4. A piezoelectric
actuator was combined with the microgripper to achieve precise grasping
motion due to its very fine positioning feature.

3
W
GI
0 +I = 1.5 mm
e
m
vl
U

a
M
e
.CI
a
a
.I
b
M
I
m
c,
-0
+ o 2 4 6 8 10 12
Input displacement (pm)
Figure 2. Total gripping distance against input displacement.
 603

0 MPa 436.828 MPa

Figure 3. Contact pressure results obtained from the FEA.

Gripper

PZT
actuator

Figure 4. The gripping device with a PZT actuator.

3.2. Tests
A movement towards the microgripper tips is transformed by the compliant
mechanism kinematics into the tip opening. Retraction of the PZT actuator
closes the microgripper tips. With the microgripper as the tool, it is possible to
grasp objects of different materials and release them. The mechanism has a
mean ampli~cationof 3.0 calculated by the average value of three sets of test
readings.
Some grasping actions were made on a 500 pm Teflon wire. Figure 5 shows
gripping of the Teflon wire in the horizontal position. The Teflon wire placed
vertically can also be gripped by the microgripper tips.
The assembly of miniaturized gear systems, typically with diameters below
2 snm, requires the use of special micro-gripping tools. The miniature gears are
604

typically of some hundreds of microns in size. Tasks of pick and place of

miniature watch gears and shafts were performed. The grasping motion was
controlled to avoid damaging the tiny teeth of the miniature gears. Figure 6
shows an isometric view of the microgripper tips gripping a shaft used for micro
gears.

Figure 5. Gripping of a Teflon wire in the horizontal position.

Figure 6 . Microgripper tips grasping a micro gear shaft.

4. A Gripper Actuated Using SMA

4.1. Prototyping
A gripper (shown in Figure 7) actuated using shape memory alloy (SMA) wire
was also fabricated and tested. The gripping force of this gripper can be
adjusted by both mechanical and electrical methods.

4.2. G r ~ p i n gForce versus Current

The relationship between gripping force and electrical current was studied. One
example of experimental results is shown in Figure 8. The figure and the value
of R2 = 0.9997 indicate that the gripping force increases linearly with the
increasing electrical current.
 605

5. Conclusion
Design, FEA, fabrication and tests of micro-grippers were conducted. The
results have proven the good performance of the micro-grippers.

Acknowledgments
The first two authors thank the Robotics Research Center of Nanyang
Technological University for the support with the piezoelectric actuator.

Figure 7. A gripping device actuated using SMA wire.

120 140 160 180 200 220 240 260

Current (a)
Figure 8. Relationship between gripping force and electrical current.

References
1. S.K. Nah and Z.W. Zhong, Sensors and Actuators A: Physical 133, 218
(2007).
2. P. Dario, M.C. Carrozza, A. Benvenuto and A. Menciassi, Journal of
Micro~echanicsand Microengineering 10,235 (2000).
606

6. E.V. Bordatchev and S.K. Nikumb, Proceedings of the International

Conference on MEMS, NAN0 and Smart Systems, Canada, 308 (2003).
7. S.H. Lee, K.C. Lee, S.S. Lee and H.S. Oh, Proceedings of The 12th
International Conference on Solid-State Sensors, Actuators and
Microsystems, Boston, 552 (2003).
8. H.Y. Chan and W.J. Li, Proceedings of the IEEE International Conference
on Robotics and Automation, Taipei, 288 (2003).
9. C.S. Pan and W.Y. Hsu, Journal of Micromechanics and Microengineering 7,
7 (1996).
10. P.B. Chu and K.S.J. Pister, Proceedings of IEEE International Conference
on Robotics anddutomation, San Diego, 820 (1994).
11. B.E. Volland, H. Heerlein and I.W. Rangelow, Microelectronic Engineering
61-62, 1015 (2002).
12. C.J Kim, A.P. Pisano and R.S. Muller, Proceedings of International
Conference on Solid-state Sensors and Actuators, San Francisco, 6 10
(1991).
13. C.J Kim, A.P. Pisano and R.S. Muller, Journal of Microelectromechanical
Systems 1,3 1 (1992).
14. C.J Kim, A.P. Pisano, R.S. Muller and M.G. Lim, Proceedings of 4th
Technical Digest of Solid-State Sensor and Actuator Workshop, Hilton
Head Island, 48 (1990).
15. P. Boggild, T.M. Hansen, K. Molhave, A. Hyldgrad, M.O. Jensen, J.
Richter, L. Montelius and F. Grey, Proceedings of the 2001 IEEE
Conference on Nanotechnology, Maui, 87 (200 1).
16. Y. Haddab, N. Challet and A. Bourjault, Proceedings of the 2000 IEEEIRSJ
International Conference on Intelligent Robots and Systems, Japan, 659
(2000).
17. M.C. Carrozza, A. Menciassi, G. Tiezzi and P. Dario, Journal of
Micromechanics and Microengineering 8, 141 (1997).
18. H. Ren and E. Gerhard, Sensors andActuators A: Physical, 58,259 (1997).
19. Z.W. Zhong and S.Y. Chan, Sensors and Actuators A: Physical 136, 335
(2007).
20. M. Kohl, E. Just, W. Pfleging and S. Miyazaki, Sensors and Actuators A:
Physical 83,208 (2000).
21. A.P. Lee, D.R. Ciarlo, P.A. Krulevitch, S . Lehew, J. Trevino and M.A.
Northrup, Proceedings of The 8th International Conference on Solid-state
Sensors and Actuators, and Eurosensors IX, Stockholm, Sweden, 368
(1995).
22. Z.W. Zhong and C.K. Yeong, Sensors and Actuators A: Physical 126, 375
(2006).
23. B. J. Pokines and E. Garcia, Smart MaterialStructures 7, 105 (1998).
 DESIGNING OF A COMMAND SHAPER USING MULTI-
OBJECTIVE PARTICLE SWARM ALGORITHM FOR
VIBRATION CONTROL OF A SINGLE-LINK FLEXIBLE
MANIPULATOR SYSTEM

M. S. ALAM ’,M. 0. TOKHI

Department of Automatic Control and Systems Engineering, University of Shefield, UK

M. A. HOSSAM
Department of Computing, University of Bradford, UK

A new command shaping method is proposed using gain and delay units to shape the
reference input in order to reduce vibration of a single-link flexible manipulator system.
The values of gain and delay elements can be derived analytically with a priori
knowledge of natural frequencies and associated damping ratios of the system, which
may not be available for complex flexible systems. Moreover, command shaping in
principle causes delay in system’s response while it reduces system vibration and in this
manner the amount of vibration reduction and the rise time conflict one another.
Assuming that, no prior information is available about the system, a new multi-objective
particle swarm optimisation (MOPSO) algorithm is applied to optimise the gain values
and the amount of delay in order to provide a wide range of solutions that trade-off these
conflicting objectives so as to satisfy design goals.

1. Introduction
Flexible manipulators are lighter, faster and less expensive than rigid ones but
they pose various challenges as compared to rigid manipulators. In order to
achieve high-speed and accurate positioning, it is necessary to control the
manipulator’s vibratory response in a cost effective manner. A good literature
review of different control strategies for flexible manipulators can be found in
(Benosman and Vey, 2004).
A feedfonvard control scheme based on input command shaping,
introduced by Singer and Seering (1990), has been applied to the control of
different types of flexible systems for vibration reduction or trajectory tracking
or occasionally both (Alam, et al., 2006; Md Zain, et al., 2006; Singh and
Singhose, 2002). The command shaping technique in practice causes delay in

607
608

system's response while it reduces vibration and the amount of reduction in

vibration and the rise time are found to be in conflict with one another.
This paper presents a new command shaping method using gain and delay
units. A new multi-objective particle swarm algorithm (MOPSO) is proposed
and applied to optimise the gain values and the amount of delay in order to
obtain a wide range of solutions, which trade-off conflicting objectives so as to
satisfy design goals.

A schematic representation of the single-link flexible manipulator system

considered in this work is shown in Figure 1, where X,OY, and XOY represent
the stationary and moving co-ordinates respectively, Z represents the applied
torque at the hub. E , I , p , T/ , I , and M , represent the Young modulus,
area moment of inertia, mass density per unit volume, cross sectional area, hub
inertia and payload of the manipulator respectively. In this study, an a l ~ i n i u m
type flexible manipulator of dimensions 9 0 0 1~ 9 . 0 0 8 ~ 3 . 2 0 0 4 ~ ~ ,
~=71x10pN/m2, I=5.253x10"m4, p=2710kg/m3, and ih=5.8598~ 10~kgm2is
considered (Tokhi and Azad, 1997).

II Accelerornete

Figure I: Schematic representation o f Figure 2: Block schematic diagram o f the

the single-link flexible manipulator experimental rig

The experimental rig, as shown in Figure 2, is equipped with a U 9 ~ 4 A T

type printed circuit motor driving the flexible manipulator. The m e a s ~ i n g
devices used to record the various responses of the manipulator are shaft
encoder, tachometer and accelerometer along the arm. The shaft encoder is used
for m e a s ~ i n gthe hub-angle of the manipulator. A precision interface circuit
 609

PCL 8 18G is used to interface the flexible manipulator system with a computer
(Pentium celeron- 500MHz).

3. The proposed command shaping for vibration control

A new command shaping method is proposed, as shown in Figure 3, using gain
and delay elements to shape the reference input. The unshaped reference signal
is passed through multiple delay units, A , and then multiplied with gain factors,
K . The shaped command is formed by summing up the delayed components.
For simplicity, the numbers of delay units and gain elements are kept the same.
The gain values are selected in such a way to give 1 when they are added
together (Singer and Seering1990). This can be shown as:

A K i =1
i=l

In order to minimise delay in system's response, the first delay unit is set to
zero, i.e., A, = 0 . Assuming that, no prior information is available about the
natural frequencies and associated damping ratios, a MOPSO algorithm is used
to find a set of solutions (delay and gain values) that trade-off between
conflicting design objectives. The proposed command shaping method is
described in Figure 4.
. Hub --
b Filter
f Flexible *"I'e
Mmi!'u'ator. Hub b
=
Filter
System Velocity -

Figure 3: Schematic diagram of Figure 4: Schematic diagram of

command shaping using gain and delay MOPSO based command shaper
elements

3.1. Objectivefunctions and goal values

Root mean squared value of residual error and rise time of hub angle response
are chosen as two objectives of the MOPSO process. The two objectives and the
corresponding goal values are shown in Table 1. Objective-1 thus chosen
corresponds to reduction in end-point acceleration response as compared to that
70% lower than the open loop response with unshaped bang-bang signal. The
610

maximum goal value of rise time (objective-2) thus chosen was 7.5% higher
than that with bang-bang unshaped response.

Objective Parameter Goal value

Obj-1 RMS value of end point acceleration 50.2
Obj-2 Rise time of hub angle response 5 1.2 sec

4. The proposed MOPSO algorithm

The dynamic equation of a modified PSO algorithm (Kennedy and Eberhart,
1995; 2001), for d -dimensional searching space is given as
Vjd = o x v i d +c1 x r a n d ( * ) x ( P j d - x i d ) + c 2 x R a n d ( * ) x b g d - x j d ) (2)

xjd = xid + Vid (3)

where c1 and c2 are positive constants, and rand(*) and Rand(*) are two
random functions in the range [0,1]; 0 is inertia weight; xid represents the i-th
particle; Vjd represents the rate of the position change (velocity) for particle i .
Pid represents the best previous position of the i-th particle and p g d represents
the best particle among all the particles in the population. These two terms, pjd
and p g d , are usually known as ‘local guide’ (or pbest)’ and ‘global guide’ (or
gbest). In case of single objective optimisation problem, pjd and p g d are
selected based on the objective function either minimum or maximum value as
far as minimisation or maximisation problem is concerned. The main challenge,
in designing a MOPSO algorithm, is to select pbest and gbest for each particle
so as to obtain a wide range of solutions that trade-off among the conflicting
objectives.

4.1. Selection method of global guide and local guide

In the proposed algorithm, a new technique is introduced that combines external
archive and non-dominated fronts of the current population in order to select
gbest for each particle. An external archive and associated control mechanism,
as used in (Coello et al. 2004), is also employed here.
For a two-objective optimisation problem, Figure 5 shows the state of the
external archive and solutions of the current particles in the objective domain.
The dark circles inside a 2-D grid structure indicate the non-dominated solutions
found so far while circles on the right represent solutions of current particles in
a 2-D objective domain and the number associated with them indicate index of
the particles in the initial population. The current solutions are sorted based on
 611

Pareto dominance and several non-dominated fronts (ND fronts) are formed as
shown in Figure 5. For each particle on ND front-1, the corresponding gbest is
selected from the external archive based on fitness sharing and roulette wheel
selection method (see Figure 5). Details of this process can be found in Coello et
al. (2004). For particles on the remaining fronts, i.e., ND front-2, 3, 4 and 5:
gbest of each particle is selected in the following way:

Figure 5: Schematic diagram for finding GBEST guide for particle in MOPSO
At first, shared fitness of each particle in the current population is
calculated based on the exact non-dominated sorting GA (NSGA) fitness
assignment scheme which was adopted by Srinivas and Deb (1994). Then, for
each particle on ND fron-2, the corresponding gbest is selected from particles
lying on the immediate lower front (better solutions), i.e., ND front-I, based on
shared fitness and roulette wheel selection method (see Figure 5). This process
continues for particles residing on the remaining ND fronts. Local guide or
pbest for each particle is selected based on Pareto-dominance (Deb, 2001).

5. Implementation
Considering the complexity of MOPSO optimisation process and required
amount of computation, the design procedure is implemented off-line. A
dynamic model of the single-link flexible manipulator based on finite element
method is utilised throughout this work. Although the flexible manipulator has
infinite number of modes with associated damping ratios, only the first few
modes appear to be dominant contributing to the flexible motion of the system.
In order to reduce vibration, mainly, at these resonance modes and to design a
command shaper having all attributes of a conventional zero vibration derivative
type (Singer and Seering,l990), both the number of gain elements and delay
units are chosen as 9. In order to minimise delay in system's response, the first
delay unit is set to zero, i.e., A, = 0 . So there are effectively 8 delay units
612

denoted as A2, A3 ,...A9 and the 9 gain values are denoted as K1, K2 )...,K 9 .
The Matlab/Simulink (The Mathworks, Inc., 2006) model (see Figure 6) is
chosen as it allows for simple construction of command shaping using gain and
delay units. Filters are used with all the three outputs with the aim to filter out
flexible motion keeping the rigid body motion intact. The cut-off frequencies of
the filters were chosen to be consistent with main vibration modes and rigid
body motion of the system. Low pass Butterworth filters were used with hub
angle and hub velocity outputs with cut-off frequency of 90Hz.

Figure 6: Simulink model of the gain-delay based command shaping control

strategy
5.1. Parameter encoding
A swarm of 20 particles having 17 elements each, i.e., 10 x 17 is created
randomly within the range of [0, +1]. The first 9 elements of each individual are
normalised and assigned to gain elements, K,, K2,...?K9 as indicated in the
Simulink model (see Figure 6). In Matlab/Similink, the delay units are usually
represented in terms of number of samples which are integer. So the remaining 8
elements of each individual are converted into integer numbers by a conversion
factor of 0.01 following a ‘round’ operation and then assigned to A2,A3, ...,A9.

6. Solutions and results

The acceleration coefficients of MOPS0 algorithm are set as c1 = c2 = 1.5 , and
inertia coefficient o is gradually decreased from 1.4 to 0.1 with generation.
The optimisation process is run for a maximum generation of 500. The
maximum number of solutions that the external archive can keep is limited to
50. In order to maintain diversity among the solutions in external archive, an
 613

adaptive grid mechanism is employed (Coello et al., 2004). The MOPSO

algorithm with a population size of 20 individuals was run on this for 500
generations. Out of a total of 10000 points evaluated, only 26 solutions were
non-dominated. The Pareto optimal set at generation 500 is shown in Figure 7
where, objective-1 is represented on the x -axis and objective-2 on the y -axis.
To validate the design approach as well as solution set, one example solution
(solutin-1) is selected on the Pareto front as shown in Figure 7. Unshaped bang-
bang input and shaped signal based on solution-1 are shown in Figure 8. The
hub angle responses for unshaped bang-bang and shaped signal based on
solution-1 are presented in Figure 9. It is clearly evident (in enlarged section)
that oscillation in the hub angle response is almost eliminated with shaped
command and it seems to settle quickly to the steady state. End-point
acceleration of the flexible manipulator due to unshaped bang-bang and shaped
signals based on solution-1 are presented in the time domain in Figure 10.

Figure 8: Bang-bang input and

shaped command (for solution-1)
Performance measures, such as, peak-to-peak amplitude, rise time, settling
time and steady-state error of the hub angle response and RMS value, norm- co
and settling time of the end-point acceleration for system response with shaped
input based on solution-1 are calculated and presented in the Table 2.
Table 2: Performance measures of different solutions
Input of open- loop Hub angle response
Control strategy Max. oscillation Rise time Settling time SS error
(peak-peak) (sec) (sec)
Bang-bang 3.1656 1.116 3.9 0
Solution-1 0.108 1.12 2.39 0

7. Conclusion
A new command shaping technique has been presented using gain and delay
units to reduce vibration of flexible systems. Considering the conflicting design
objectives, such as, reduction of vibration and speed of response, a new
MOPSO algorithm has been successfully applied to derive solutions based on
614

two objectives, satisfying the design goals, known as Pareto-optimal set, which
describes the trade-off among conflicting objectives. The Pareto front yields a
set of candidate solutions, from which the desired one is picked under different
trade-off conditions. All the performance measures give a good indication of
reduction of vibration in system's response compared to unshaped response.

-46 i " 3imrSeR s

Figure 9: Hub angle response for bang- Figure 10: End-point acceleration
bang input and shaped command (for due to bang-bang input and
solution-1) shaped command (for solution-1)
References
Alam, MS, Tokhi, MO., Siddique, MNH. and Hossain, MA. (2006). Selection and
Designing of command shaper using multi-objective genetic algorithms for vibration
control of a single-link flexible manipulator, Proceedings of the qhInternational
Conference on Climbing and Walking Robots and the Support technologies for
Mobile Machines, 11-14 Sept., Brussels, Belgium.
Benosman, M. and Vey, L. (2004). Control of flexible manipulators: A survey, Robotica,
22,535-545.
Coello, CAC., Pulido, GT. and Lechuga, MS. (2004). Handling Multiple Objectives with
Particle Swarm Optimization, IEEE Transactions on Evolutionary Computation,
8(3), 256279.
Deb, K. (2001). Multi-objective optimization using evolutionary algorithms. New York;
Chichester: Wiley.
Kennedy, J. and Eberhart, R. (1995). Particle swarm optimization, Proc. of IEEE
International Conference on Neural Networks, IV, 1942-1948, Perth, Australia,
Kennedy, J. and Eberhart, R. (200 1). Swarm Intelligence, Morgan Kaufmann Publishers.
MATLAB Reference Guide, The Math Works, Inc., 2006.
Md Zain, MZ., Tokhi, MO. and Mohamed, Z. (2006). Hybrid learning control schemes
with input shaping of a flexible manipulator system, Mechatronics, 16(3-4), 209-219.
Singer, NC. and Seering, WP. (1990). Preshaping command inputs to reduce system
vibration, J. Dynamic Systems, Measurement and Control, 112(l), 7 6 8 2 .
Singh, T. and Singhose, W. (2002). Tutorial on input shapingltime delay control of
maneuvering flexible structures, Proceedings of 2002 American control conference,
Omnipress, Madison, 1717-173 1.
Srinivas, N. and Deb, K. (1994). Multiobjective Optimization Using Nondominated
Sorting in Genetic Algorithms, Evolutionary Computation, 2(3), 221-248.
Tokhi, MO. and Azad, AKM. (1997). Design and development of an experimental
flexible manipulator system, Robotica, 15(Part 3), 283-292.
 DETECTING SOUND SOURCES WITH THE HUMANOID
ROBOT RH-1

PAVEL STAROVEROV, RICARDO MARTINEZ, DMITRY KAYNOV, MARIO

ARBULU, LUIS CABAS, CARLOS BALAGUER
Robotics Lab, Department of System Engineering and Automation, University Carlos III
of Madrid, Av. Universidad 30, Leganks 2891 I (MADRID), Spain

Rh-l is a humanoid robot under development at the University Carlos I11 of Madrid.
Along with other interaction types, the robot is capable of recognizing voice commands.
This article describes the sound localization system that has been added recently. It is
independent of the voice recognition system and uses the information From several
microphones attached to the robot’s head that is further processed in order to estimate the
sound source direction. Results are presented and discussed and the ideas how to improve
the performance are suggested.

1. Introduction
The ability to detect the direction of sound sources can be a u s e l l addition to
the interaction system of a humanoid robot or another type of service robots. In
case if the speaking person is not facing the robot, the robot could turn to the
speaker or just turn its head, thus making the communication more natural and
pleasant for the person.
Rh-1 is a humanoid robot developed at the University Carlos I11 of Madnd.
Rh-1 has 21 degrees of freedom (+2 more degrees of freedom of the built-in
camera), is 130 cm tall and weighs in about 45 kg (Fig. 1).
From the point of view of interaction, the robot is capable of speech
recognition, speech synthesis, face detection I recognition and the so-called “T-
shirt based” recognition. Those abilities were discussed in [16]. There must be
another paper somewhere in the present proceedings describing the newly added
gesture (posture) recognition system based on shirt shapes.
The objective of the work was to develop an easy-to-implement sound
source detection system with a good correspondence between the invested time
and effort on one hand and with an acceptable performance on the other hand.
The present article is organized as follows: in Sec. 2 some related works are
discussed, Sec. 3 describes the development of the system, in Sec. 4 the
experimental results are presented and discussed. Section 5 concludes the paper.

615
616

2. Related works
One of the first works in the field of acoustic perception for robots made in
1995 at MIT involved the robots Cog and Kismet 111. SIG and his successor
SIG2 (developed at the University of Kyoto since 2000) have auditory systems
based on binaural hearing that use 2 pairs of microphones (one pair is used to
cancel the noises produced by the motors) [2,3]. ROBITA robot developed at
Waseda University can follow a conversation between two persons using his 2
microphones [4]. Lately there has been a trend to use more than 2 microphones.
For example, one of the most advanced humanoid robots HW-2 developed by
Kawada Industries and AIST has 8 microphones [5], Sony QEUO small sized
humanoid robot has 7 of them [6]. Finally, there is the famous Honda ASIMO
humanoid robot that is able to localize sound sources, making difference
between voices and other sounds or noises 173. Concerning more recent works,
there is a robot called Wakamam developed by Mitsubishi that uses microphone
array for sound source localization [S].
Speaking of the algorithms used, there are several signal processing
techniques that are used in robotics and in other applications as well, among
these are: energy based methods [9], methods based on the time difference of
the signals received by a pair of microphones like “Cross-power Spectrum
 617

Phase" [10]-[14] and fmally there is the "Sum and Delay Beamformer"
technique that is used to process the signals from a 128 microphone array [ 151.

3. ~ e v e l o ~ ~of
enthe
t system
The objective was to develop a system that would let the robot detect the
position of the speakmg person relative to the robot. This implies using several
microphones, converting their signals into the digital form and processing the
data. First of all, let us take a look at the hardware components that the system is
ma
dT
e
o
.
hf
esc
he
mec
v
a
i
wt
ht
o
es
fy
s
et
m s
p
i
e
rse
n
et
n
d
Fi
g
2i
..

I
1
i
I
I
I I
I
i
K
1
I
I I
I
I i '----- ! I
I I
I
I
1 t
r I
I I
r I
I I
I
1
I L
I I
I
I t
i I
I
I
I
I
1
I
I
L II - - -- ---- II -I
Fig.2. Hardware of the sound localization system.
618

There are 4 Knowles Acoustics MD9755USZ- 1 unidirectional

microphones that are located in the lower part of the robot’s head in the
horizontal plane, one in the front, one in the back and two at both sides. Those
microphones feed the signals into the custom made preamp board with. It has
the shape of a flat donut (it was necessary to design the board that way to fit it
into the robot’s head) and can amplify up to 6 microphones, though currently
only 4 of them are used. The gain of each channel can be adjusted manually.
Some of the voltages needed by the board are provided by the PC/104 power
supply (that also powers all the computer components of the robot), but the most
noise-sensitive components are powered by 3 standard AA batteries. It helps to
reduce the amount of noise picked up by the preamps. The board is connected to
the RTD SDM7540 data acquisition board that performs the A/D conversion (12
bits, 22 kHz sampling rate). The CPU board Digital Logic MSM855 is running
Windows X P operating system.
Now let us take a look at the software and the algorithms used. A driver
written in C reads the data from the acquisition board and passes it to a Matlab
program for further processing. First, we apply a band-pass filter that
accentuates the voice frequencies and eliminates the unwanted noises and DC
offset in the microphone data. Next, we calculate the RMS value of N samples
of each microphone’s readings {.,J using the following formula:

During the experiments a value of N equal to 1000 samples was used. After
that a threshold function is applied to eliminate the background noises.
Threshold value equal to 0.015 mV Rh4S proved to be a reasonable one for the
normal conditions of our laboratory that has several functioning PCs and an air
conditioning system. Then the readings from all the microphones are compared
and the result is assigned in each timeframe of N samples, 0 if all the values are
below the threshold, and the number of the microphone (1-4) that has the
loudest signal in the contrary case. Finally, the number of hits of each
microphone is compared during a larger period of time (5 seconds for the most
of the test results), and the winning microphone is determined. This result can
be passed to other programs running on the robot that can use it in conjunction
with the voice recognition and image processing systems. In Fig.3 you may see
a window with the results of the program’s execution during one of the
experiments, there are 4 graphs with the signal amplitude for each microphone
 619

a nh
a
g
dt
e
a
s
lr
t
ph
sh
ow
nh
g
i
t
ewn
n
ig
i
m c
o
r
i
pho
n
n
ee
i
a
mci
hta
e
rm
f T
e
h
.e
n
i
e
r
a
f
sa
si
l
ul
p
tl
osho
w

If* I

1 I I
1
I I I

I I I
3

,%

&
I

I
1-7
! \
\-i
I , I
1-1, {
* L1 rn II

N L ‘ N E F DF THE VICR3PHONE 3
Fig.3. An example of the results.
The method used is very simple and only provides 90 degrees of accuracy.
However, that might be enough in case of our application. The main reason why
this method was selected is because it is easy and fast to implement. Let us see
the experimental results to see whether the performance is acceptable for our
application or not.

4. Results and discussion

There have been three series of experiments, in each one of them the speaker
was repeating the same sound (or phrase) three times in front of every
microphone. The first test was with the person producing 2 handclaps, the
second one was with a 1-second-long phrase “Robot, I’m here”, and the thrd
one was with a longer, 3-second-long phase “Robot, I’m here, do you know
where I am?” Every test was repeated at 1 m distance from the robot and then at
2 m distance. The result was considered positive, if the number of the
microphone returned by the program coincided with the speaker’s position and
negative in the contrary case. The total number of tests was 72. The robot was
placed at 2.5 m distance from the walls, nearly in the center of the room. The
results are presented in Table 1.
620

Table 1. Results of the tests.

Type of the sound Results at 1 m distance Results at 2 m distance

Two handclaps 7 positive / 5 negative 2 positive / 10 negative
Short phrase 9 positive / 3 negative 7 positive / 5 negative
Long phrase 1 1 positive / 1 negative 8 positive / 4 negative
The general impression is that the performance is acceptable only at a short
distance (1 m) and only with phrases, with handclaps the system presented a
high number of misses which is unsatisfactory. The performance of the system
with long phrases pronounced at a short distance, which is supposed to be the
most frequent case for our robot, may be considered quite good, with more than
90% of the results being correct.
The performance of the system with handclaps and at greater distances is
worse because of the fact that under these conditions the reverberations from the
walls become quite prominent, masking the direct sound. The algorithm used
currently is quite straightforward and makes no effort to cancel those echoes.
The solution may be to use a more robust, more advanced time-based algorithm
like Cross-power Spectrum Phase described in [ l l ] and [12]. Also, if the sound
source direction detection will be used during the robot’s motion, the sounds
generated by the motors should be canceled like suggested in [2] and [3].
To summarize, it can be said that the performance of the developed system
is acceptable for out type of application, however, there are ways to improve it.

5. Conclusion
The present article suggests a sound source direction detection system for a
humanoid robot. Sound source direction detection may be a valuable addition to
the interaction system of any service robot. The proposed system is based on 4
microphones and a simple energy-based algorithm. The development of the
system was explained in detail, along with the algorithms used. Some
experimental results were presented and discussed. The performance of the
system is considered acceptable for our needs. Several ways to improve the
performance of the system are suggested. Future works may include
implementation of a more robust time-based algorithm that would also cancel
reverberations and noises from the motors.
 621

Acknowledgments
We would like to thank the members of the Robotics Lab of University Carlos
I11 of Madrid for their cooperation and suggestions. The present work was
supported by CICYT (Comision Interministerial de Ciencia y Tecnologia).

References
1. R. Brooks, C. Breazeal, M. Marjanovie, B. Scassellati, and M. Williamson.
“The Cog project: Building a humanoid robot,” in Computation for
Metaphors, Analogy, and Agents, C. Nehaniv, Ed. Springer-Verlag, 1999,
pp. 52-87.
2. Kazuhiro Nakadai et al., “Active Audition for Humanoid”, Kitano
Symbiotic Systems Project, ERATO, Japan Science and Technology COT.
3. K. Nakadai, H. G. Okuno, and H. Kitano, “Real-time sound source
localization and separation for robot audition”, Proc. IEEE ICSLP, 2002,
pp. 193-196.
4. Yosuke M. et al., Perceptual Computing Group, Waseda university,
“Modelling of Conversational Strategy for the Robot Participating in the
Group Conversation”, Eurospeech 200 1.
5. Isao Hara, Futoshi Asano et. al., “Robust speech interface based on audio
and video information fusion for humanoid HRP-2” Proc. ICIRS, Sendai,
September 2004, pp. 2404-2410.
6. Fujita M. et al., Intelligent Dynamics Laboratory, Sony Corporation,
“Autonomous behaviour control arquitecture of entertainment humanoid
robot SDR-4X”, International Conference on Intelligent Robots and
Systems, 2003.
7. Sakagami Y. et al. “The intelligent ASIMO: System overview and
integration” Proceedings of the 2002 IEEE/RSJ Intl. Conference on
Intelligent Robots and Systems, EPFL, Lausanne, Switzerland,pp.2478-83
8. Tomonaka T. et al. “Computer Vision Technologies for Home-use Robot
“Wakamaru”, Mitsubishi Heavy Industries, Ltd. Technical Review Vol. 42 No.
1 (Feb. 2005)
9. T. Ikeda et al., “Framework of Distributed Audition”, Proc. 13th IEEE
International Workshop on Robot and Human Interactive Communication,
Sept. 2004, Japan, pp. 77 - 82.
10. M. Brandstein, A Framework for Speech Source Localization Using Sensor
Arrays, Ph.D. thesis, Brown University, 1995.
11. M. Omologo, P. Svaizer, “Use of the Cross-Power Spectrum Phase in
acoustic event localization”, IRST Trento, Technical Report #9303- 13.
12. M. Omologo, P. Svaizer, “Acoustic Event Localization using a Crosspower
Spectrum Phase based Technique”, Proc. ICASSP, Adelaide 1994, pp.
11273-11276.
622

13. M. Omologo, P. Svaizer, “Talker Localization and Speech Enhancement in

a Noisy Environment using a Microphone Array based Acquisition
System”, Proceedings Eurospeech, Berlin, September 1993, pp. 605-609.
14. Amir A. Handzel, P. S. Krishnaprasad, “Biomimetic Sound-Source
Localization”, IEEE Sensors Journal, V01.2, No.6, December 2002, pp.
607-6 16.
15. Y. Tamai et al., “Real-Time 2 dimensional Sound Source Localization by
128-Channel Huge Microphone Array”, Proc. IWRHIC, Okayama, Sep
2004, pp. 65-70.
16. P.Staroverov; M.Arbul6; L.M.Cabas; D.Kaynov; C.PCrez; C.Balaguer. “A
Voice Controlled Image Recognition System.” Proc. of the 9th International
Conference on Climbing and Walking Robots. Brussels. Belgium. Sep,
2006.
 IN SEARCH OF PRINCIPLES OF ODOUR SOURCE
LOCALISATION

E. E. KADAR
University of Portsmouth, Department of Psychology, King Henry Building, King Henry
I Street, Portsmouth, PO1 2DX UK.

G. S . VIRK
Massey University, School of Engineering and Technology, Wellington, New Zealand.

C . LYTRIDIS
Department of Electronic and Computer Engineering, Anglesea Building, Anglesea
Road, Portsmouth, PO1 3DJ UK.

Arguably, chemical sensing is perhaps the most fundamental sensory modality in animals.
This modality is far simpler than most other modalities (vision, hearing, heat sensing,
electric and magnetic sensors) and yet navigation in a chemical diffusion field is still not
well understood. Biological studies have already demonstrated the use of various search
methods (e.g., chemotaxis and biased random walk) by various animals, but robotics
research could also provide new ways to investigate principles of olfactory-based search
skills [1],[2],[3]. The primary aim of the present paper is to demonstrate that similar
navigational principles can be observed in various animals ranging from invertebrates to
mammals but the use of these principles can lead to complex search strategies.
Importantly, search strategies derived from these biological principles tend to result in
similar patterns in both simulation studies and robot experiments.

1. INTRODUCTION
After more than a century of research on animal behaviour and numerous
recent studies with robots, the principles of navigational strategies in chemical
fields are still not well understood. For instance, early behaviourist experiments
tried to explain behaviour in terms of reactions to stimuli and introduced various
reaction including klinolunesis, klinotaxis, tropotaxis, etc. In general, too many
reactions were needed to explain search behaviour for all observable conditions.
To simplify theoretical account of search patterns, researchers attempted to use
or rely on some intuitive principles. For instance, Fraenkel and Gunn [4]
emphasized that chemoreceptors are not direction receptors and they cannot be
used to orientate the animal to a gradient in a chemical field. Later, these

623
624

different animal reactions to chemicals seem to have been forgotten and all
search strategies in chemical fields were, in an overly simplistic way, called
chemotaxis [5], [6]. During the past few decades, researchers were strongly
influenced by the cognitive revolution and tended to created artificial search
algorithms by often ignoring the fact that chemical search behaviour is highly
influenced by environmental constraints such as the nature of the diffusion field
(highly instable and especially noisy far from the chemical source) as well as the
biological constraints such as the limitations of sensors (number of sensors,
slow sensory response, problem of sensory adaptation). To further complicate
matters, robotic researchers tend to confound their search models by using other
sensory modalities without being clear on the use of principles and limitations
of chemical searching. These problems highlight the importance of having a
clear and consistent terminology and use them in relation to fundamental
principles in experiment tests.
Motivated by these ideas, we set out to achieve a better understanding of
the principles of search strategies in chemical diffusion fields. First we
investigated biological strategies and tried to implement them in simulation
studies. Biased random walk (BRW) strategies are shown to be robust and yet
more efficient than chemotaxis in unstable and noisy chemical fields [7]. In
addition to locating the static point odour source in unstable chemical fields,
BRW has also been assessed for moving targets [8], odour trails [9] and
turbulent plumes [lo], [ll].
Most of our previous work has focused on testing navigational strategies in
simulation studies. In agreement with Webb [l], we believe that robotics
research provides new ways to investigate biological strategies. In particular, the
principles of olfactory navigational strategies can be distilled from various
search patterns (based on biological and simulation studies) and systematically
investigated on robotic platforms to overcome some of the limitations of
simulation studies (e.g., difficulties in modelling odour diffusion and dispersion
processes when a moving agent introduces additional noise in the chemical
field).
Although the poor quality of artificial sensors does constrain research and
simulation as well as robot-based experiments are insufficient to find accurate
models for biological search performance, we argue in the present paper that
investigation of animal search patterns, simulation studies and robot
experiments provide sufficient evidence that only a few principles are needed to
explain complex olfactory search patterns observed in natural settings.
 625

2. BASIC PRINCIPLES OF NAVIGATIONAL STRATEGIES

Traditionally, in the literature chemical search is often called chemotaxis.
Recognizing that taxis behaviour require two sensors, it becomes clear,
however, that chemotaxis should only be used for search strategies that rely on
directional information to orientate the agent towards the direction of a chemical
gradient. Given the fact that chemosensors are not direction sensitive [4] and
can only use local information (in contrast to visual and auditory sensors) a
single sensory reading cannot provide sufficient information on the direction of
a gradient [ 101. Although many species have bisensory chemical receptor organs
(antennae, nostrils), quite often they have to rely on only one of these sensors
(e.g., when one of the sensors is impaired). These facts led us to turn our
attention to unisensory search mechanisms.
In our early simulation studies, we have investigated unisensory BRW as an
alternative search strategy to chemotaxis [7], [S]. Also, we argued that most bi-
sensory species are probably rely on unisensory strategies even though they use
both unisensory and bisensory (chemotaxis) strategies. Thus, one of the
fundamental principles is that a chemical sensor is not direction sensitive but if
it is combined with partly random locomotion it could be robust and effective to
follow gradient information under various conditions. Another important
principle we learned from our investigations is that bisensory capability is
important for many species to provide some redundancy for unisensory search
but at the same time two sensors could enhance efficiency of unisensory search
under a variety of conditions (e.g., high concentration in the neighbourhood of
target etc.). The present study provides some examples from animal research as
well as simulation and robot studies to support the validity of these principles
and demonstrate how they give rise to complex search patterns.

2.1. Animal behaviour in chemical search

Search patterns of small worms as well as higher order species exhibit some
fundamental similarities. Typically, at far range from target, search patterns are
erratic and wildly meandering. Nevertheless, animals gradually drift close to the
target. As the target is approached, the movement patterns are getting less
meandering and often either getting straight or mildly sinusoid around an axis
pointing toward the target. In [12], Koehler has reported movement patterns of
Planaria lugubris clearly showing these generic characteristics (see Figure 1a).
In a pilot study with a blind cat, we have also observed these patterns (see
Figure lb). Interestingly, the overall characteristics of these search patterns can
626

be observed in bisensory chemotaxis search as well as unisensory search

performance.

target

(a) (b)
Figure 1: (a) Schematic drawing of typical search patterns in Koehler’s (1932) study with planarians.
(b) Search pattern of a blind cat toward a dish with cat-food in it. In both animals, the early stage of
search at a distance from target is characterized by meandering patterns and gradual drifts closer to
target. At a close to target range, the movement becomes nearly straight with some pendulation of
the head.

2.2. Chemotaxis in simulation and robot research

Chemotaxis is based on the detection of a concentration difference between
two chemical sensors and a steering mechanism toward the direction of hgher
intensity while moving forward at a constant speed. In principle, chemotaxis-
based navigation could result in smooth movement trajectories in smooth
chemical fields without much environmental noise. Practically, however, this
strategy is not sufficient and, as we have shown [7], often leads to failure due to
the intrinsically noisy and highly unstable chemical field conditions. Robotic
implementation create additional problem by the fact that robot movement
fkther increases the noise in the diffusion field. Thus, the bisensory orientation
becomes ineffective under noisy condition and at low concentration level (i.e.
far from target range). However, moving in a high concentration andor stable
smooth diffusion field the bisensory system can facilitate search patterns that are
driving the agent toward the target with a nearly smooth trajectory with some
addition meandering factors.
Figure 2 shows successful search patterns by chemotaxis from simulation
and robot search. Two main characteristics.of these patters can be observed: 1)
 627

At a far range, the search performance is not promising because it is

meandering but the agent still drifts closer to the target. 2) Close to the target,
the search pattern becomes less erratic and often results in an oscillatory
movement whle approaching toward the target.

.i-
1.4r

1.2 -

as -
as--
s* a4--

0.2- -
0- *

0.2- -
a 4
1.5
L I
-1
,
0 5
I
0
I
0.5
,
1
I
1.5
X&

Figure 2: Chemotaxis-based search patterns of robot trials and simulations demonstrating that
search performance is meandering in a noisy field at far range but it becomes smooth and slightly
meandering driving the agent toward the target if the field is stable [13]

2.3. Biased random walk

Lytndis et al. argued in [13] that, in principle, unisensory search could be
based on chemotactic principles because unisensory temporal sampling of the
chemical concentration field could be interpreted as a method of compensation
for the absence of spatially separated sensors in a unisensory organism. But this
‘chemotaxis mimicking’ strategy would require a memory for registering spatial
locations with associated intensity values of the chemical field to calculate a
new movement direction. The use of this complicated computational mechanism
is unlikely in primitive organisms and it may not be necessary in higher-order
species either.
And indeed, studies suggest that bacteria as well as other higher order
species use a much simpler unisensory strategy consisting of straight runs with
occasional directional changes (tumbling) [5]. Alternating these two modes of
motion, the animal generates polygon shaped trajectories. It has been shown that
both the run lengths and directional changes can be described by Poisson
distributions. The characteristics of movement patterns depend on the
concentration distribution in the search environment. In a homogeneous field,
there is no gradient information and the resulting motion is a purely random
walk. In an inhomogeneous environment, the distribution h c t i o n of the run
lengths is continuously changing since the mean of the distribution depends on
628

the detected concentration gradient (e.g., intensity difference between two

different locations). When the concentration level increases, the frequency of
directional changes decreases, and therefore the forward runs tend to be longer.
The result of this strategy is an overall drift toward the source of the diffusion
field. To put it differently, the random walk becomes a biased random walk
process. This is a simple but efficient strategy using a fast adaptation process to
the local field conditions instead of relying on memory of previously measured
intensities at various locations in the search process. During the search process,
the moving organism can adjust its sensory organs to the local concentration
level and can detect relative increases or decreases in field intensities when it
drifts into another region.
A similar approach has been developed into a BRW algorithm at a
macroscopic scale. Its use in artificial agents was first demonstrated in noisy
Gaussian fields [7]; and this work was extended to dynamic fields [lo]. The
BRW strategy in odour source localisation has been shown to be a robust
method under a variety of conditions, but its efficiency can be improved [7], [8].
7 im
I I I I I I I I I I I I I

:j2

Figure 3: Search patterns of robot trials and simulationsusing biased random walking
Although BRW is successful in driving the searching agent close to the
target from regions of low concentration level, BRW is less efficient in the
neighbourhood of the target because it tends to generate longer runs which may
lead to ‘mistakes’ such as overshooting or passing by the target. We have tried
to remedy t h s shortcoming based on the observation that search patterns in
animals get less meandering as they get closer to the target. This can be
achieved by halting the agent at a certain high intensity threshold near the target.
Similar result could also be achieved by reducing the range of random
directional changes as well as reducing the run length bias above an intensity
threshold typical close to the target. In both cases, search patterns remain
unisensory.
 629

Figure 3 shows a few search trials from simulation as well as robotic

search. Typically, search patters are irregular especially far from the target
because the field intensity values are smaller and the noise makes the field less
stable and more difficult to use to detect gradient. Despite this difficulty, the
agent gradually drifts toward the target and as it gets closer to the target where
field intensity tends to be higher even though the noise is also present. Thus, it
is more likely that high concentration field provides better opportunities to
detect gradient information on the direction of target during a straight run.

2.4. Combined strategy

With intact bisensory systems, animals can benefit from both the unisensory
as well as the bisensory search strategies and animal search patterns seem to
confirm t h s hypothesis. This is why we have developed combined search
stategies in our simulation studies and robot experiments [13]. Far from target
with low field intensity and noisy environment, unisensory BRW search patterns
are robust and efficient to drift the agent closer to the target. When the agent
reaches a higher intensity threshold characteristic of the region close to the
target, the unisensory BRW strategy can still be applied with a small
modification. The bisensory system can be used to introduce a directional bias
in the otherwise random directional changes. The extent of this bias can be
dependent on the concentration level of the chemical field. Figure 4 shows a
trial with a combined strategy exhibiting similar movement patterns shown in
previous examples (See Figures 1-3).

Figure 4: Search patterns of robot trials using the combined strategy

630

3. CONCLUSION
Recent experiments in robotics are often based on complex algorithms and
the performance is assessed by success in artificial environment. In the past
decade, we followed a different strategy. We tried to identify fundamental
principles of biological search strategies in relation to the intrinsic properties
(e.g., weak especially at a far range, lack of stability, and fast changing
characteristics, etc.) of chemical diffusion fields [ 131, [14]. The present study
provided a brief but hopefully convincing highlight of a decade long systematic
research programme. We have started our research by observing biological
search strategies in various species ranging from bacteria to cats. Then we
formulated some hypothesis on fundamental principles of search strategies
(unisensory versus bisensory in relation to field conditions) and tested them in
simulation studies. Finally, we implemented these strategies in robotic platforms
and our experiments resulted in similar search patterns observed in animals. In
sum, our starting point was observation of animal search patterns to derive
fundamental principles of biological chemical search strategies and we
concluded our research by implementing these principles in ‘artificial search
algorithms’ on robotic platforms to create search patterns similar to the ones
created by animals.
We have to finish this paper, however, with a cautionary note. In
biologically inspired robotics research as well as other biological modelling
experiments, scientists have to be careful and remain clear about the distinction
between simulation and mimicking. Although in this paper we used the word
simulation and robotic implementation of simulation, these studies actually
cannot be considered proper simulations. They are simply just mimiclung
biological strategies. We should be allowed to use the word “simulation”
properly if the actual robot sensory detection and further processing would be
similar to the processes animals use in their remarkable search performance.
Nevertheless, this cautionary note does not diminish the value of our research
because we believe the search performance should be based on the interaction
field of the organidagent and its environment when a performance is based on
principles of search strategies in a natural environment.

References
1. Webb, B. (2000). What does robotics offer animal behaviour? Animal
Behaviour, 60 545-558.
2. Grasso, F. W., Consi, T. R., Mountain, D. C., & Atema, J. (2000).
Biomimetic robot lobster performs chemo-orientation in turbulence using a
 63 1

pair of spatially separated sensors: Progress and challenges. Robotics and

Autonomous Systems, 30(1-2), 115-131.
3. Lytridis C, Virk GS and Kadar EE (2005). Search performance of a multi-
robot team in odour source localisation, Proceedings of 8th International
Conference on Climbing and Walking Robotics (CLAWAR 2005), London,
3-15 September.
4. Fraenkel G. S. and Gunn D. L. (1961). The orientation of animals, Dover,
New York.
5. Berg, H. C., & Brown, D. A. (1972). Chemotaxis in Escherichia coli
analysed by three-dimensional tracking. Nature, 239(5374), 50CL504.
6. Beer, R. D. (1990). Intelligence as adaptive behaviour: An experiment in
computational neuroethology. Perspectives in artificial intelligence no. 6.
Boston: Academic Press.
7. Kadar, E. E., & Virk, G. S. (1998). Field theory based navigation for
autonomous mobile machines, Proceedings of the IFAC Workshop on
Intelligent Components for Vehicles (ICV '98), Seville, Spain, 23-24
March.
8. Kadar, E. E., & Virk, G. S. (1998). Field theory based navigation towards a
moving target. Advanced Robotics: Beyond 2000: 29th International
Sympo-sium on Robotics, Birmingham, England, 27 April - 1 May.
9. Virk, G. S., & Kadar, E. E. (2000). Trail following navigational strategies.
Proceedings of the 3rd International Conference on Climbing and Walking
Robots (CLAWAR 2000), Madrid, Spain, 2-4 October.
10. Lytridis, C., Virk, G. S., Rebour, Y., & Kadar, E. E. (2002). Odour-based
navigational strategies for mobile agents. Adaptive Behaviour, 9(3-4), 171-
187.
11. Mielle, P., Marquis, F., & Latrasse, C. (2000). Electronic noses: specify or
disappear. Sensors and Actuators B: Chemical, 69(3), 287-294.
12. Koehler, 0. (1932). Sinnesphisiologie der Susswasserplanarian. Zeitung
Vergl. Physiologie, 16,606-756.
13. Lytndis, C., Kadar, E. E., and Virk, G. S . (2006) A systematic approach to
the problem of odour source localisation. Autonomous Robots, vol. 20,3,
pp. 261-276.
14. Kadar E. E. (1996). A field theoretic approach to the perceptual control of
action. PhD Dissertation. University of Connecticut, USA.
 GA TUNED CLOSED-LOOP CONTROL OF SPRING BRAKE
ORTHOSIS

M SAIFUL HUQ, RASHA MASSOUD, M SHAFIUL ALAM AND M 0 TOKHI

University of Sheffield, Sheffield Sl 3JD, United Kingdom

Abstract: Spring brake orthosis (SBO) is a kind of hybrid orthosis system (HOS) which
generates the swing phase of paraplegic gait by employing a spring at the knee joint to
store energy during the knee extension through quadriceps stimulation, which is then
released to produce knee flexion. The acceptance of any HOS and hence its degree of
success depends largely on its performance in generating an acceptable gait trajectories.
In this paper the performances of PID and Fuzzy logic controller (FLC) is studied in
generating the swing phase in an SBO equipped human leg. A leg model of an average
sized (183 cm height) is developed for simulation purposes. The leg model includes
segmental dynamics, passive joint viscoelastic properties, an electrically stimulated knee
extensor quadriceps muscle and an SBO. An optimal knee joint trajectory to be used as
the reference knee joint trajectory is also derived. All the controller parameters for both
PID and FLC are tuned using genetic algorithms (GAS) for minimum tracking error.

1. Introduction

Complete or partial loss of ability to walk or stand due to lower limb paralysis is
a very common as well as drastic result of thoracic level spinal cord injury
(SCI). The loss of lower limb function and inability to walk or stand
significantly reduce the quality of life of disabled individuals and carry with
them psychological and physiological effects. The spectrum of problems that
may interfere with locomotory performance after and SCI includes hyperactivity
of spinal reflex (muscle spasticity) [l], alternation in the muscle activation
patterns, including weaknesses and difficulty in coping with weight bearing,
balance and gait speed [2].
The electrical stimulation of nervous system below the level of the spinal
lesion can produce powerful muscle contractions and thus can be used to
generate primitive movements. This is termed as ‘functional electrical
stimulation’ (FES) and is used to obtain a functional, useful movement by
evoking artificial contraction of the muscles deprived of nervous control. The
aims of the restoration varies with the ambition of the researchers and range
from assistance with wheelchair transfers to the ability to stand up and sit down,

632
 633

to take few steps, to walking for some distance. Most work has concentrated on
the correction of foot drop in hemiplegia and on the restoration of standing and
walking in paraplegia.
The main challehge in FES assisted movement arises from the fact that the
artificially stimulated muscle fatigues very quickly because of the 'reversed
recruitment order' of the artificially stimulated motoneurons. The consequences
are twofold: (a) limited duration of the FES assisted movement, especially
standing and walking, (b) drastic changes in the actuator (muscle) properties
lead to poor movement control. One of the major approaches to overcome these
limitations is to reduce the use of active muscle, where possible, through the use
of passive braces and is called hybrid orthosis systems (HOS) [3]. Several
research groups have tested many HOS in the laboratory [4]. The spring brake
orthosis is a kind of HOS, which generates the swing phase of paraplegic gait.
Some open- and closed-loop FES control strategies that were designed and
tested are described in [ 5 ] . The open-loop FES control performance was
however found unsatisfactory for accurate movement control due to existing
parameter variations (e.g., muscle fatigue), inherent time-variance, time delay,
and strong nonlinearities present in the neuromuscular-skeletal system (plant).
Closed-loop control is thus necessary for accurate control of movement, and
adaptive closed-loop strategies are obviously the most suitable candidates to
tackle the FES control problem.

1.1. The Spring Brake Orthosis Concept

The quadriceps muscles, when artificially stimulated, could produce much more
torque than is required just to extend the knee. During knee extension of swing
phase, SBO exploits this feature of the quadriceps through partially storing FES
generated quadriceps force as potential energy in a torsion spring attached to the
knee joint. A brake is then employed to maintain the knee extension without any
muscle contraction. Then knee flexion is achieved by releasing the brake and
letting the spring to return to its resting position (approx 70' - 80'). The hip
flexion is simultaneously produced as a result of consequent shift in the centre
of mass (CoM) of the overall leg segment during this knee flexion and is
maintained throughout the required duration by applying a brakehatchet at the
hip joint. This results in a hybrid orthosis combining electrical stimulation of
quadriceps muscle, spring and brake at the knee joint and brakehatchet at the
hip joint, with the activation of each of them at appropriate instant and for
appropriate period [6]. Bearing in mind that an acceptable gait can be achieved
634

with a locked ankle [7], the ankle joint is kept fixed with an ankle-foot-orthosis
(AFO).

2. Methods

2.1. Model
A forward dynamic planner (sagittal) human leg consisting of segmental
dynamics [8], passive properties at the joints [9], electrically stimulated muscle
model of quadriceps [lo] and SBO was modelled through a combination of
visualNastran@(VN) software and Sirnulink@.The SBO design parameters are
taken from a previous study which employed Genetic algorithms (GAS) to find
the best parameters.

2.2. Control of SBO

Since the SBO aided swing phase involves no control at the hip joint, its control
in generating swing phase refers to the control of knee joint trajectory (mainly
during knee extension) through quadriceps stimulation. Both PID and PD type
Fuzzy logic controller (FLC) are tested in the closed-loop form. All the
parameters of the controller are optimised using GAS for carefully chosen
objective functions. Cycle-to-cycle control [ 111 is also investigated and
optimised due to its increasing popularity and evident superiority over closed-
loop control in FES control [12]. Once again, a GA is employed to search for
the best parameters for the cycle-to-cycle control strategy.

2.2.1. Reference Trajectory

Attention is paid to derive the best trajectory for the closed-loop controllers.
Two factors affected the choice of trajectory, (a) maximizing the energy
efficiency following suggestion from [ 131, so as to increase energy efficiency as
well as reduce muscle fatigue, @) its degree of ability to drive the knee joint all
the way to full knee extension. Since cycle-to-cycle control takes care of only
the end-of-phase orientation of the concerned limb joint rather than taking it
through a predefined trajectory, the joint trajectory thus traversed is of
significant importance in terms of achieving full knee extension. Bearing this in
mind, the knee joint trajectory traversed under an optimised cycle-to-cycle
control strategy is used as reference trajectory in closed loop control strategy.
 635

2.2.2. Controller Tuning with GAS

The controller parameters are initialised using Ziegler-Nichols Method (for PID)
and heuristic method (for FLC). Several search operations are performed to find
the best values for the controller parameters using GAS.
GAS, first proposed by Holland in 1975 [14], are search and optimisation
techniques inspired by two biological principles of natural evolution, viz. the
process of "natural selection" and the mechanics of "natural genetics". Contrary
to regular search algorithms, GAS manipulate not just one potential solution to a
problem, but a collection of potential solutions, called a population.
Real world problems, in most cases, involve multiple objectives to be
achieved simultaneously. Due to the conflicting nature of the objectives, it is
often difficult to find a single optimal solution for a problem, and makes more
sense to seek for a set of nondominated perato-optimal solutions for which an
improvement in one of the objectives will lead to degradation in one or more of
the remaining objectives. GAS are a suitable technique for multiobjective
optimisation due to there population based nature which enables them to support
multiple solutions concurrently. One of the first approaches to utilise the
concept of pareto optimality in GAS was Fonseca and Fleming's multiobjective
GAS(MOGA) [ 151.

3. Results

3.1. Closed-loop control (Reference Trajectory Derived from Passive

Oscillation)
The reference trajectory for the knee, being inspired by [13], is derived from
simple passive oscillation of the shank in absence of any stimulation
whatsoever. Both PID controller and FLC are used in an attempt to drive the
knee joint through the given trajectory and implement the swing phase. All the
controller parameters for both PID and FLC are optimised using a GA. For the
PID controller, only 3 parameters, viz. P, I and D are optimised whereas the
with FLC, it is 73 parameters, including the weightings of 25 rules (5x5rule
base) and parameters associated with triangular member functions (MF). For
each of the controllers the GA was run for 100 generations with 50 individuals
in each generation, thereby evaluating 5000 points for each of the parameters.
The integral of the 'time-weighted squared error' (ITSE) is defined as the
objective function to be minimized. This time weighting in the objective
function is to emphasize the tracking error nearer to the knee extension so as to
prioritise the necessity of full knee extension.
636

Figure 1 shows the decay of the objective functions with generation while
Figure 2 shows the resultant controlled trajectories achieved with the optimised
controller parameters. Although it’s obvious from both the figures that the FLC
perfoms much better than the PID controller in terms of input tracking, both of
them are unacceptable in practice due to their inability to produce full knee
extension. This suggests that, given this particular trajectory both these
controllers are incapable of producing full knee extension. To put it in a
different way, with the given plant (human leg) with all its limiting
characteristics, this particular trajectory is probably unsuitable for this particular
purpose.

Generation
Figure 1. GA optimisation of PID controller and FLC - decay of objective function with generation
 637

Time (secf

Figure 2. Knee and hip joint trajectories - reference trajectory derived from plain passive oscillation,
actual controlled trajectories under PID control and FLC.

3.2. Cycle-to-cycleControl

Figure 3. Optimised cycle-to-cycle control

The model, at this stage, is subjected to cycle-to-cycle control with a fixed pulse
with of 220 ms, while the burst duration as well was its position in the time
scale is optimised formulating a minimization problem using a MOGA for two
638

objectives, (a) square of the deviation of the produced knee extension from 0”
(objective 1)’ (b) time-integral of the active torque (from quadriceps) (objective
2). Minimizing ‘objective 1’ helps producing an exact 0’ knee extension, while
it contributes towards less muscle fatigue in case of ‘objective 2’ [16].
From the obtained pareto-optimal solution set, the one with a minimum
value of ‘objective 1’ (near 0” knee extension) is chosen and the result is shown
in Figure 3. Also presented in the same plot the optimised stimulus burst and
resultant active torque.

3.3. Closed-loop Control with Reference Trajectory Derived from Cycle-

to-cycle Control

Figure 4. PID and FL control of knee joint trajectory for SBO swing phase
The optimised cycle-to-cycle control strategy with a successhl end-of-phase
knee extension results in a unique knee joints trajectory. This uniqueness is
mainly characterized by the minimum time-integral of the active torque from the
stimulated muscle. This unique trajectory is then used as the input reference for
the closed-loop controllers of section 3.1 with the controller parameters tuned
again in the same way.
The success in introducing the new reference trajectory is very much
obvious in Figure 4, which presents the resultant knee joint trajectory under
FLC and PID control along with the reference trajectory. The result, besides
asserting a high degree of suitability of the given trajectory, also reveals the
 639

variation in the degree of suitability amongst arbitrary trajectories for FES

applications.
Hip joint trajectory is found to be satisfactory (25"+4") in all the cases, and
that's why it is not presented in all the results.

3.4. Comparison between the Controllers

0' 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9
Time (sec)

Figure 5. Accumulation (integral) of active muscle torque during swing phase with different
contro11ers
In FES applications, muscle fatigue is always an inevitable pitfall and hence
always desirable to keep it as minimum as possible. Figure 5 is a brief
representation of performances of the controllers investigated so far, in terms of
fatigue immunity. The figure shows time-integral of the active quadriceps
torque generated under each control strategy which is related to the energy
consumption of the muscle and thus to the fatigue.
The two curves indicating least amount of time-integral of active torque
correspond to the initial closed-loop control approach with modified passive
oscillation as reference trajectory. Although they seem to produce potentially
less fatigue, they are of no significance due to their inability to produce full
knee extension. Among the other 3 curves cycle-to-cycle control seems to
accumulate least active muscle torque. FLC appears to accumulate less active
torque between the rest 2 curves of which both are closed-loop.
640

Control % Normal Excursion

,
Cycle-to-cycle z 83%

PID 86.57%

FLC 91.1%

(a) (b)
Figure 6 . (a) Knee joint trajectories with the same controllers of section 3.3 under fatigue condition
(50% of normal quadriceps torque output), (b) Table showing percentage drop in the excursion of
the knee under fatigue condition.
Although apparently the open-loop cycle-to-cycle control would produce
least fatigue, the closed-loop controllers, due to their obvious nature are more
robust against plant variations, e.g. due to fatigue. This is revealed in the result
in Figure 6 (a) and (b), where Figure 6 (a) shows the trajectories with the same
controllers, but active torque halved, simulating fatigue condition, while the
table in Figure 6 (b) quantitatively presents the same result in terms of knee
joint excursion as a percentage of the excursion in normal (fatigueless)
condition. It is obvious that both the closed-loop controllers are more robust
than the cycle-to-cycle controller with the FLC at the top.

4. Concluding Remarks
This work primarily concentrates on closed-loop tracing control of SBO assisted
swing phase with a view to benefit for the advantages of feedback tracking
control. One of the main attractions of cycle-to-cycle control lies in achieving
the necessary end-of-phase orientation required for successful fimctional
movement without the need of any trajectory. Being discrete-time in nature, it
also suffers from the lack of real time online supervision of states [16]. In
principle, the current work combines the advantages of both the control
strategies. A successful practical implementation of any FES control must
include adaptation mechanism for the controller parameter(s), especially to cope
with fatigue related changes. Such an optimised controller, as the one in this
work, can serve as an initialised part of an adaptive controller through the
addition of an online adaptation mechanism.

Reference
1 . Eltorai, I. and R. Montroy, Muscle release in the management of
spasticity in spinal cord injury. Paraplegia, 1990. 28(7): p. 433-40.
 641

2. Waters, R.L., et al., Recovery following complete paraplegia. Arch

Phys Med Rehabil, 1992. 73(9): p. 784-9.
3. Goldfarb, M., et al., Preliminary evaluation of a controlled-brake
orthosis for FES-aided gait. IEEE Trans Neural Syst Rehabil Eng,
2003. ll(3): p. 241-8.
4. Durfee, W.K. and A. Rivard, Design and simulation of a pneumatic,
stored-energy, hybrid orthosis for gait restoration. J Biomech Eng,
2005. 127(6): p. 1014-9.
5. Crago, P.E., et al., New control strategies for neuroprosthetic systems.
J Rehabil Res Dev, 1996. 33(2): p. 158-72.
6. Gharooni, S., B. Heller, and M.O. Tokhi, A new hybrid spring brake
orthosis for controlling hip and knee flexion in the swing phase. IEEE
Trans Neural Syst Rehabil Eng, 2001.9( 1): p. 106-7.
7. Goldfarb, M. and W.K. Durfee, Design of a controlled-brake orthosis
for FES-aidedgait. IEEE Trans Rehabil Eng, 1996. 4(1): p. 13-24.
8. Winter, D.A., Biomechanics and motor control of human movement.
2nd ed. 1990, New York ; Chichester: Wiley. xvi, 277p ill 25cm.
9. Amankwah, K., R.J. Triolo, and R. Kirsch, Effects of spinal cord injury
on lower-limb passive joint moments revealed through a nonlinear
viscoelastic model. J Rehabil Res Dev, 2004. 41(1): p. 15-32.
10. Riener, R. and T. Fuhr, Patient-driven control of FES-supported
standing up: a simulation study. IEEE Trans Rehabil Eng, 1998. 6(2):
p. 113-24.
11. Franken, H.M., et al., Cycle-to-cycle control of swing phase of
paraplegic gait induced by surface electrical stimulation. Med Biol
Eng Comput, 1995.33(3 Spec No): p. 440-5 1.
12. Arifin, A., T. Watanabe, and N. Hoshimiya, Design of Fuzzy
Controller of the Cycle-to-Cycle Control for Swing Phase of
Hemiplegic Gait Induced by FES. Ieice Transactions on Information
and Systems E Series D, 2006. 89: p. 1525-1533.
13. Williamson, M.M. Exploiting natural dynamics in robot control. in
Cybernetics and systems research. 1998. Vienna: Austrian Society for
Cybernetic Studies.
14. Holland, J.H., Adaptation in Natural and Artificial Systems. 1975, Ann
Arbor: The University of Michigan Press.
15. Fonseca, C.M. and P.J. Fleming. Genetic algorithms for multiobjective
optimisation: Formulation, discussion and generalization. in Fifth
International Conference on Genetic Algorithms. 1993. Morgan
Kaufmann, San Mateo, CA.
16. Veltink, P.H., Control of FES-induced cyclical movements of the lower
leg. Med Biol Eng Comput, 1991. 29(6): p. NS8-12.
 Hidden Markov Model based Fuzzy Controller
for Flexible-link Manipulator
M.N.H. SIDDIQUE'
'School of Computing and Intelligent Systems, University of Ulster

M.A. HOSSAIN'
'Department of Computing, University of Bradford

M.S. ALAM3AND M.O. TOKH14

3.4Departmentof Automatic Control and Systems Engineering, University of S h g j e l d

A major problem with fuzzy rule-based systems is that with an increasing number of inputs
and linguistic variables, the possible number of rules for the system increases exponentially.
Unfortunately, there is no systematic approach to learning of rule-base of fuzzy logic (FLC)
controller if there is no control expert available, then it must be constructed from the
controlled environment or a suitable data set should be available. The adaptive neuro-fuzzy
inference system (ANFIS) proposed by Roger Jang, which reduces the development time
involved in constructing the rule-base requires a set of input-output data. The problem is now
how to cope with developing an FLC where apriori information such as a set of input-output
behaviour or expert knowledge is not directly available. The hidden Markov model (HMM) is
a probabilistic finite-state machine used in finding structures in sequential data. A rule-base of
an FLC can be compared to a finite state machine which can produce a sequence of output
MFs. Therefore, the main interest of this research lies in finding a functional mapping from a
rule-base of FLC to a hidden Markov model (HMM) and train the HMM using the available
data source. The developed controller is then applied to a flexible-link manipulator to verify
the performance of the methodology.

1. Introduction
In most existing applications, the fuzzy rules are generated by an expert in
the area, especially for the control problems with only a few inputs. As is
well recognized, rule acquisition has been and continues to be regarded as a
bottleneck for implementation of any kind of rule-based systems. A major
problem with fuzzy rule-based systems is that with an increasing number of
inputs and linguistic variables, the possible number of rules for the system
increases exponentially, e.g. for a three-inputs and single-output system with

642
 643

7 primary sets for each input will require n x n x n = 7 x 7 x 7 = 343 rules.

This huge number of rules makes it difficult for experts to define a complete
set of rules and associated membership functions for a sufficient level of
system performance. The difficulty arises if there is no control expert
available for constructing the rule-base, then it must be constructed from the
controlled environment or a suitable data set should be available. Another
well-known problem with fuzzy rule-based systems is that the processing of
such a rule-base is time consuming. Consequently, most of the time is spent
in calculating the control output using center of gravity method of
defuzzification, which in some applications can cause slow system response.
The problem of such defuzzification methods has been eliminated by the use
of Sugeno-type fuzzy systems, where each consequent fuzzy set is replaced
by a linear function [I]. This imposes a further set of consequence
parameters to be estimated. Current neural-fuzzy systems are mainly
Sugeno-type fuzzy systems with the rule-base replaced by a neural network.
The most widely used neuro-fuzzy system is the adaptive neuro-fuzzy
inference system (ANFIS) proposed by Roger Jang, which reduces the
development time involved in constructing the rule-base [2]. The learning of
the parameters of the neuro-fuzzy system requires a set of input-output data
and uses several passes of backpropagation and/or LMS algorithm to
estimate the antecedent parameters of MFs and consequent parameters of the
system [ 1-31. The problem is now how to cope with developing a fuzzy logic
controller (FLC) for a flexible-link manipulator, which is a highly nonlinear
system [4,5]and apriori information such as a set of input-output behaviour
or expert knowledge for constructing rule-base is not directly available. The
hidden Markov model (HMM) is a probabilistic finite-state machine used in
finding structures in sequential data [lo]. A rule-base of an FLC can be
compared to a finite state machine which can produce a sequence of output
MFs. The control input is thus generated by defuzzifying output MFs using
usual methods.
The objective of this research is to minimize the processing time
required for the control input by avoiding Mamdani-type inferencing using
rule-base and centre of gravity defuzzification procedure. Therefore, the
main interest of this research lies in finding a conceptual mapping from a
rule-base of FLC to a hidden Markov model (HMM) and train the HMM
644

using the available data source. The developed controller is then applied to a
flexible-link manipulator to verify the performance of the methodology.

2. HMM-Fuzzy controller
The basic principles of the HMM-fuzzy controller is that for a given rule-
base as defined by a global linguistic association constrained by fuzzy sets, a
functional mapping from the fuzzy logic-based controller to the HMM-based
approach is established. In other words, the inference mechanism (or
approximate reasoning mechanism) of a rule-based FLC is implemented by a
HMM. A block diagram of the proposed HMM-fuzzy control system is
shown in Figure 1.

Trained off-line using fuzzy granules

...........................
~

Figure 1: Block diagram of the HMM-Fuzzy Controller.

2.1 Fuzzification
A membership function (MF), such as large, medium or small, is typically a
fuzzy set. The linguistic variable is defined by the following Gaussian
membership function (MF).

Where rn denotes the central value and CT denotes the spread of a Gaussian
MF.
 645

2.2 HMM as Rule-base Mapping

Unfortunately, there is no systematic approach to learning of rule-base of fuzzy
controller. Efforts have been made to automate the construction of rule-bases in
various ways using neural networks and genetic algorithms [6-81. In most of the
cases, either the rule-base is fixed and the parameters of the membership functions
are adjusted or membership functions are fixed and the rule-base is optimised by
genetic algorithms [S]. The rule-base of a Mamdani-type fuzzy controller with two
inputs (mostly error and change of error) and a single output (torque input) consists
of the following rules.
R: IF x,is Ai AND x,is B j THEN U is c k (2)
where X,& error, X,&change of error, U control input, Ai , i = 1,2,. ,N ,-
- -
B J. , j = 1,2, * ,M are the input linguistic variables and c k , k = 1,2, * ,L are
the output linguistic variables, N, M and L are the maximum number of linguistic
variables (primary fuzzy sets defined by membership functions). The rule-base of
the Mamdani-type fuzzy controller is shown in Tablel, where c,is defined as
c, Ec k = (c, c, -
, ,* , c, } and each c, is to be found from available
information granules.
Table 1: Rule-base R
Input X,

The rule-base R in Table 1 can be shown as a mapping @ from input fuzzy sets to
output fuzzy sets defined by the equation ( 2 ) . The mapping @ can be visualized as
mapping of information granules from inputs to output pictorially as shown in
Figure 2. More descriptively, the particular interest in this research is the
development of possibility functions p,,,,, p B, from informatio granules (Ai},
f
{Bj}[I 11, which will compute the probability /unction P ( 0 I il' from the MHH
model 2 , i.e. HMM will produce {c,
} such that it satisfies the requirements of the
mapping @ in (3).
a) : { A i , B j } + C, (3)

D
646

The functional mapping : R + @ can be represented by HMM, defined by

x ={TR,E,z} (4)
where TR is the transition probability matrix, E is the emission matrix and n is the

e
initial probabil' . TR = [a,]is an N x Nmatrix and the elements should satisfy
a, 2 0 and a..= 1. Such an HMM is shown in Fi ure 3, where each a, is
comparable tZFA, (X,),pEJ
defined such that it satisfies the condition crkA,r
(x~)}.T h e h n c ion A, ( x l ) , p (x2)} is so
(x, ,pBJ(x2)7= 1. In other
words, the inference mechanism of a rule-bas4 Mamdani-type FLC is implemented
by the HMM in Figure 3. Once the HMM is trained, it is then inserted in the control
loop for on-line operation.

2.3 Defuzzification
The most time-consuming operation in Mamdani-type FLC is the defuzzification
procedure using center of gravity method. To minimize the time required for
defuzzification, Gaussian MFs are used in the consequent part. Since the Gaussian
MFs are symmetric and invertible, defizzified value is calculated using equation (5)
while ignoring the left half of the MFs, shaded area, shown in Figure 4.

Information granule: X2

PL-n

Figure 2: Rule-base as a Mapping

 647

BN
Figure 3: HMM as a rule-base mapping.

Figure 4: Defuzzification of output fhzzy sets.

648

Where mdenotes the central value, and o denotes the spread of a Gaussian
linguistic variables. A suitable support set for output fuzzy sets ci
is to be defined
so that the aggregated crisp value does not fall outside the universe of discourse.
That is, the support comprises those elements of y of the universe such that
pc( y ) > 0 . In this case, support set is d fined as
I
support(cj = [y pc, (Y)2 p, j (6)
Where ps is chosen arbitraril for each output fuzzy set
1
P, = PC, (A I mi 5 x < mi+]
It is straightforward and easy to find the value of pus
f
and defined as
(7)
empirically for all output fuzzy
sets, which satisfies equation (7).

3. Experimental rig
The experimental rig constituting the flexible manipulator system consists of two
main parts: a flexible arm and measuring devices. The flexible arm contains a
flexible link driven by a printed armature motor at the hub. The measuring devices
are shaft encoder, tachometer, accelerometer and strain gauges along the length of
the arm. The shaft encoder, tachometer and accelerometer are essentially utilised in
this work. The experimental flexible-link manipulator is shown in Figure 5.
The flexible arm consists of an aluminium-type beam. The outputs of the sensors
as well as a voltage proportional to the current applied to the motor are fed to a
computer through a signal conditioning circuit and an anti-aliasing filter for analysis
and calculation of the control signal. Physical parameters of the flexible arm are
given in Table 2.

4. Experimental results
The HMM-fuzzy controller is constructed with 6 fuzzy sets (linguistic variables) for
two inputs and 3 fuzzy sets for output and hence the HMM has 3 states as shown in
Figure 3. HMM is trained using the control torque collected from a bang-bang
controller applied to the flexible-link manipulator, quantised in fuzzy interval NS, Z
and PS, labelled them as C,,Cz, and C3 respectively. The transition matrix TR and
emission matrix E are initialised arbitrarily using simple heuristic which is the main
advantage of training HMM. 800 data points were used to train HMM to an error
goal of 0.001. The HMM-FLC produced control torque is shown in Figure 6. The
endpoint acceleration is shown in Figure 7.
 649

Table 2: Physical parameters.

Parameter Value

Length 960.0 ~IJI

Width 19.008m

Thickness 3.2004 mm
2710
Mass density/ kgm”
Figure 5: Flexible-link manipulator unit volume

2.8

2.6

2.4

2.2

1.8

1.6

1.4

--____;
_ _ _ _ _ - j __---_;
__-___;
______;_ _ _ _ _ _
1.2

1
100
I
I
200
I- . --
300
I I

400
______

500
Time units in 0.015 sec
~

I
I
I
I
600
I1
700
-
800

Figure 6: Control torque produced by HMM-FLC

650

2 4 6 8 10 2
lime in seci

Figure 7: Endpoint acceleration.

5. Conclusion
This paper has been presented an investigation to minimize the processing time
required for the control input by avoiding Mamdani-type inferencing using rule-base
and centre of gravity defizzification procedure. A conceptual mapping from a rule-
base of FLC to train the HMM using the available data source has been developed to
demonstrate the merits of the proposed system. The developed controller was
applied to a flexible-link manipulator to verify the performance of the methodology
through a set of experiments. It is noted that the proposed controller performed
reasonable well and very much similar to the traditional FLC.

References
1. T. Takagi and M. Sugeno (1 985). ‘‘Fuzzy identification of systems and its applications to
modeling and control”, ZEEE Transaction on System, Man and Cybernetics, Vol. 15,
1985, pp. 116 -132.
2. J.4. Roger Jang, (1993). “ANFIS: Adaptive-network-based fuzzy inference system”, ZEEE
Transaction on Systems, Man and Cybernetics, Vol. 23, No. 3, 1993, pp. 665-685.
 65 1

3. D.A. Linkens and J Nie (1994). “Backpropagation Neural Network based Fuzzy Controller
with a Self-learning Teacher”, International Journal of Control, vol. 60, No. 1, 1994, pp.
17-39.
4. S. Yurkovich, F.E. Pacheo and A.P. Tzes, (1989). “On-line Frequency Domain Information
for Control of a Flexible-link Robot with varying Payload”, IEEE Transaction on
Automatic Control, Vol. 34, No. 12, 1989, pp. 1300-1304.
5. M.N.H. Siddique, (2002). “Intelligent Control of Flexible-link Manipulator Systems”, PhD
Thesis, Department of Automatic Control and Systems Engineering, The University of
Sheffield, England, UK.
6. Lin, C.-T. and Lee, C.S.G. (1995). “A neural fuzzy control system with structure and
parameter learning”, Fuzzy Sets and Systems, Vol. 70, pp. 183-212.
7. Nauck, D and Kruse, R (1993). “A fuzzy neural network learning fuzzy control rules and
membership functions by fuzzy error backpropagation”, Proceeding of IEEE International
Conference on Neural Networks, pp. 1022-1027.
8. Nauck, D and Kruse, R. (1996). “Designing neuro-fuzzy systems through
backpropagation.” In W.Pedrycz, ed. Fuzzy Modelling: Paradigms and Practice, pp. 203-
228, Kluwer, Boston.
9. H Bourlard and C. Wellekens, (1990). “Links between Hidden Markov Models and
Multilayered Perceptrons”, IEEE Transaction on Pattern Analysis and Machine
Intelligence, vol. 12, pp. 1167-1178, 1990.
10. L. R. Rabiner (1989). “A Tutorial on Hidden Markov Models and Selected Applications in
Speech Recognition”, Proceedings of the IEEE, Vol. 77, No. 2, pp. 257-286.
11. L.A. Zadeh (1978). “Fuzzy Sets as a Basis for a Theory of Possibility”, Fuzzy Sets and
Systems, Vol. 1, pp. 3-28.
HIL/SIL BY DEVELOPMENT OF SIX-LEGGED ROBOT SLAIR2

DZHANTIMIROV', PAL IS^, SCHMUCKER', TELESH~,ZAVGORODNIY~

'Department Virtual Engineering, Fraunhofer Institute for Factory Operation and
Automation, Sandtorstrasse 22, 39106 Magdeburg, Germany
'Institute for Electrical Energy Systems, University of Magdeburg, Universitatsplatz 2,
P.B. 4120,39106 Magdeburg, Germany

The development process of CLAWAR is a complex task. The Hardware- and Software-
in-the-loop frameworks, which are used by development of new 22 DoF six-legged
robot SLAIR2 are presented. A novel universal real-time communication bridge system
is introduced and test results are presented. The ability to real-time communication and
service tasks related control of the robot is discussed.
Keywords: Embedded and Real Time Systems, Modelling and Simulation Languages,
Automatic Code Generation.

1. Introduction
The control system of the full or part of autonomous legged robot is almost
controlled by embedded system. Commonly the embedded systems are designed
to control complex plants such as engines, satellites, vehicles, spacecrafts, and of
course CLAWAR. They generally require a high level of complexity within the
embedded system to manage the complexity of the plant under control.
Development and test of complex real-time embedded systems consists of
many steps from modelling and simulation of the plant till the implementation of
the source code in the real hardware. Hybrid techniques, that are used
increasingly, are the Hardware-in-the-Loop (HiL) and Software-in-the-Loop
(SiL) simulations (see Fig.1). In our work we use these methods within the
development and testing process of six-legged robot.
The embedded system (i.e. real electronic control device or real
mechatronical component) within a HiL-framework is connected to
corresponding HiL-simulator (which emulates real system response) via its 10s
building a closed loop. The system to be controlled (i.e. legged robot) is then
simulated to test the correct performance of the control system (i.e. drive
control). The inputs of control system under test are stimulated with the model-
based sensor outputs. To close the control loop is the reaction of the control
system outputs (i.e. motor voltage) guided back into robot model.

652
 653

controller object
h
\

PC

(MATLAWSimulink, etc.)

II - test

...................................
object E III-operation

Figure 1. Common scheme of the HiUSiL structure of the mechatronical system.

In the pure simulation phase a virtual control system is connected to a

virtual object. Both software systems communicate via pure software interface.
In the hybrid simulation phase for such communication a hardware interface
needed. The development process is through communication inconsistence more
complex and needs more effort to transfer of results [ 5 ] . And moreover with the
change of communication interface is even more effort to expect.
The paper presents the developed six-legged robot “SLAIRT [3] as well as
the environment, which was used in the pure simulation phase and the hybrid
HiL/SiL phase by development and test separate robot components. Moreover
the flexible communication bridge for real-time communication between the
virtualheal control system and the virtualheal robot is presented as well.

2. Modular Six-Legged Robot “SLAIR2”

Figure 2 represents the multi-legged robot with articulated body “SLAIR2”
that has been developed at the Fraunhofer Institute for Factory Operation Otto-
von-Guericke University both from Magdeburg in Germany [4]. The robot
mechanics, sensor system and control system make it possible to maintain the
additional flexibility in the body, to measure and control the support reactions as
well as to control and forecast the robot motion stability.
654

2.1. Mechanics and Sensor Systems

The rdbot construction consists of n = 3 modular segments (shoulders)
linked to each other through two DOF joints and 6 legs. Each shoulder includes
one articulated body segment linked with two 3-DOF-insectomorphic legs. It is
possible to extend the construction to the case of n > 3 shoulders. The main
mechanical parameters are shown in the Table 1. The robot drives are
servomotors, with maximal power P,, = 4.5 W, with potentiometer with
angular range qpon=+90" and gears with ratio iCEAR= 251 and
efficiency vGmR = 85% .

Table 1 . The main mechanical parameters of the robot

leg lengths total: ILEG = 300 mm
thigh: 11 = 120 mm
shank: 12 = 180 mm

The sensor system of the robot consists of components that are standard for
mobile robots and that make it possible to achieve autonomous robot functions
in an environment. It includes:
22 potentiometers and 22 current sensors (installed in each robot joint),
6 three-component force sensors (mounted in each leg's shank),
2-axis gyroscopic sensor (located in body), and
 655

further it will be equipped with near navigation sensor system (stereoscopic

camera).
In accordance with the requirements on measurement and control of the
support reactions, the developed force sensor consists of two parts: the core
measuring lateral components of support reaction, and the elastic parallelogram
module for measurement of the longitudinal component. The sensor is designed
for loadings up to FcoNTAcT= 50N; interference between channels does not
exceed 1%.

2.2. Control System und Control Structure

The hierarchically organized modular control structure (Fig.3a,3b,3c) is
completely located on PC-side that implement additionally the interaction with
user and produces the control signal for robot drives as well as monitors all
actuators and sensors of the robot. The robot-side is implemented by fast and
flexible FPGA, includes the hardware abstraction layer (HAL,) for drive and
sensors. The real-time connection between two parts is made via proposed netX@
[ 11 based communication bridge described below. All this provides flexibility
and simplifies the development of control algorithms. The control system can
also be extended for the additional shoulders in exactly the same manner as the
mechanical structure.

ACTION LEVEL : REFERENCIES FOR LOCOMOTION TASKS

Support Polygon Step Circle Kurs Direction Task Specified Task Specified
Geometrie & Rotation & Rotation Robot Impedance Body Config
1 1 1 1 1
ACTION LEVEL : REFERENCIES FOR MANIPULATION TASKS
Body Trunslation Task Specified
& Orientution Tool
1
Figure 3a. The hierarchically organized robot control system: action level
656

PRIMITIVE LEVEL : TRAJECTOR.IES GENERATION

SERVO LEVEL: TRAJECTORIES CONTROL 13 MONITORING

Task

wcs
AC7

Figure 3c. The hierarchically organized robot control system: servo level

3. NetX Communication Bridge and its implementation in

Matlab/Simulink Environment
The communication bridge emulates a router, which connects the control
system with all the components of the control object (i.e. legged robot). These
components can be connected to the bridge via diverse bus systems (Ethernet,
CAN, UART, SPI etc). The robot components in such system can also be
represented in virtual environment. Through a common physical interface is the
virtual or real control system isolated from the object, and therefore the
development effort of communication interface to the legged robot is reduced.
The output signals from the control system (reference signals) are de-
multiplexed and sent to corresponding actuators with the help of the
communication bridge. On the other side the outputs from the robot (actual
 657

sensor signals) are multiplexed by bridge and sent to the control system. These
routing tasks are entirely performed by the communication bridges. The physical
implementation has been done using -9-based netX - SoC [ 11, which is able
to communicate with all well known fieldbus and real-time Ethernet systems.
Figure 4 represents the proposed flexible communication system based on
netX communication processors.

full duplex full duplex full duplex

I
. ....._
.................... -.............
...............................................
Figure 4. Flexible communication system based on netX processors.

The parameterisation of the communication bridge takes place in control

program of embedded ARM processors. This program can be compiled with
conventional tools, but in order to simplify the reconfiguration of the bridge
within simulation framework the programming via RTW and GUI in
Matlab/Simulink was selected.
For these purposes a special software stack (s.Fig.5) was developed and the
communication bridge was integrated into Matlab/Simulink framework (s. Fig.6)
under Real-Time Workshop [ 2 ] .

i
ethernet definer

initetho \
Figure 5. Software Ethernet communication stack implemented in Matlab/Simulink.
658

Figure 6. Integration of netX based communication system into M a ~ a b / S i m u l i ~ .

This framework allows convenient development and HiLlSiL tests of the

control system and generation of the application code for the target netX
processors under ~atlab/Simulinkenvironment only (s.Fig.7).

Figure 7. HiLlSiL smcture by development of six-legged robot SLAIR2

4. Test of the Nets based Corn Bridge in Red Six-Legged Robot

For the execution of the communication tests a soft and a hardware

configuration have been suggested, which represents the SiL - application. The
 659

hardware part of the test platform consists of a host PC and two netX units. The
units are interconnected via serial communication chain.
The communication is implemented in full duplex mode, whereby
simultaneous sending and receiving of the packets are possible. The task of the
host PC is in the periodic sending of request and the collecting of all response
packages in the same communication cycle with all netX units (s.Fig.8)
.........................

I.....................................................
kster Slave I ,.......k.....s.....t...e
~
.....r................................................... Slave slave...._I

......................................................................................... .................................................... _,
j

!........Master
......................................................................Stave
........... Stave j ;........Master
............................................. Slave ;

Figure 8. Data flow within communicationcycle between Host-PC and netX chain.

The packets are led to and from the second netX unit “B” over the first netX
unit “A”. In order to reduce the turn-around time of the those Ethernet frames,
whose goal MAC address does not agree with the own MAC address of the
current unit, the first netX unit “A” functions in the so-called switch mode.
To test purposes the response time of netX units is interesting, which
corresponds to the performance of the program generated by the RTW, as well
as the entire cycle time for the operation of the netX chain with the Simulink
model, running at the host PC. To check the performance of the whole system
the forward and backward communication flow with distributed control units
have been used (s.Fig.9).

Host-PC
netXA
netX B
Switch-delay

Host-PC
netXA
netX 0
SwitchMay Cycle timedifferenca

Figure 9. Forward and backward communicationsequence with netX ‘A’and netX ‘B’
660

The timing measurement is accomplished in certain places over GPIO exits

and observed thus at the oscilloscope (s.Fig.10). The measurement shows the
incoming and processing times of the request packages for both netX units.
. . . . . . . . . . . . . . , , . . . ! . . . . ! . . . . ! . , . .
. . . i""+""" . . . .
. .:
. :131P% a : $ : : : :
. . . 7, . . . .
. . . . . 't . . . : . . . . . . . . . . . . . . . . . . .
. . .
. . .
:
. . . . . . . . . :.- . . .
. . .
. . . . .

~ . . . . .
. . . . , . . . . ,. . . . . . . . . ,. . . . .. i . .. . . , . . . . ., . . . . . $ . ..
/
.
, / , . . .
.
2.00V Ch2 2.00V M50.0ps
Figure 10. Packet incoming and processing time for netX 'Aand netX 'B'

The test results (table 2 and 3) show the maximum cycle rate of about 3 kHz
for the presented test system.

0000000 0000066pp 0000200 0000293 200 227 293

1016545 1016638 1016747 1016863 202 225 318
2016530 2016593 2016727 2016819 197 226 289
3016509 3016570 3016706 3016794 197 224 285
4016503 4016589 4016701 4016826 198 237 323
5016486 5016570 5016684 5016802 198 232 316
Average runtime: T~round/ T ~ r ~ /Tround
nd 3 - 199 229 304

0000000 0000100 0000228 0000297 228 197 297

1011693 1011764 1011926 1011958 233 194 265
201 1657 201 1729 201 1883 201 1927 226 198 270
3011631 3011720 3011850 3011910 219 190 279
4011627 4011697 4011860 4011888 233 191 261
5011598 5011666 5011833 5011853 235 187 255
271

The cycle times for separate netX units let estimate the switch turn-around
time of approximately 15-18 ps, whereby the pure transmission time of 100 byte
 661

package via 100 Mbit/s Ethernet is about 10 ps. This indicates the fact that the
passing package is completely buffered by the netX switch.
The source code generated by the RTW for the netX units is not optimal
concerning to the response time. Therefore the additional manual optimization of
this source code allows the reduction of the cycle time from about 270 ps down
to 90 ps.

References
1. http://www.hilscher.comHilscher GmbH, Hattersheim, Germany
2. http://www.mathworks.com/vroducts/rtwembedded/
3. http://www.uni-magdeburg.de/ieathobotslab
4. F.Palis, V.Rusin, USchmucker, ASchneider, Y. Zavgorodniy. Legged
Robot with Articulated Body in Locomotion over Complex Terrain. 7th Int.
Conference on CLAWAR, 22-24 September 2004, Madrid.
5 . F. Kanehiro et.al. Distributed Control System of Humanoid Robots based on
Real-time Ethernet. IEEE/RSJ Int. Conference on Intelligent Robots and
Systems, 9-15 October 2006, Beijing
 IMPROVING POWER TO WEIGHT RATIO OF
PNEUMATICALLY POWERED LEGGED ROBOTS

G. MCLATCHEY, J. BILLINGSLEY
Faculty of Engineering and Surveying, University of Southern Queensland, West St,
Toowoomba, QLD 4350, Australia

Legged Robots require actuators with a high power to weight ratio. Although pneumatic
actuators do not perform well in this regard, they have other attractive characteristics
which are useful in Legged Robots. This paper describes a mechanical solution for
significantly improving the payload capacity of a robot powered with pneumatic
cylinders, Robug IV,and reports on the theoretical design and experimental outcomes.
Keywords: Pneumatic cylinders, legged robots.

1. State of Pneumatically Powered Legged Robot Research.

The supporting legs of legged robots form part of multiple closed kinematic
chains in which antagonistic forces pose a problem. Such a system is much
better suited for force control, which can accommodate the over-constrained
system [l], [2], [3], than position control where small errors can increase
antagonistic forces [4], [5], [6].
Fluid based actuators such as hydraulics and pneumatics are ideally suited
for force control, however they are less popular than electric motors because
their natural compliance makes position control difficult and they are low in
power density [2], [7].
Control of pneumatic cylinders has been advanced as much as possible by
the authors’ work in Nested Loop Control [4], however, the inherent problem of
low power to weight ratio remains. The most promising solution is to re-direct
some of the torque available from the actuator to lift a leg into pushing the leg
down to hold the body up. This has been implemented in Robot V but not
investigated [8].
Here, such a solution is designed, implemented and reported for the robot
Robug IV. Also, the maximum increase of performance and the extent of the
benefits of such a solution are discussed.

662
 663

2. Design
Figure la) shows the original Hip joint for Robug IV. Note the Hip actuator
between points F and E and the Knee actuator (which is not considered further)
between points G and I. The Hip on a Robug IV leg is the major load-bearing
joint and must balance the largest force acting on the robot, which is its weight
due to gravity. On uneven terrain the Abductor and Knee joints must provide
higher torque to balance Robug’s weight, but a strategy of re-directing more
torque in a particular direction is unsuitable since similar torque is required in
both directions of joint movement.
The simplest way of re-directing torque is to use a spring element. The
basic design of a spring inclusion is shown in Figure lb). The original link
between B and D is replaced with a plate that provides a spring mounting point
C. A tension spring stretches between points C and I, and provides a positive
torque around the Hip axis at 1. Please note that an identical spring and plate is
provided on the other side of the Hip.

Figure 1. a) Robug IV Hip Joint, b) With Spring, c) Geometric Lengths and Angles Identified.

2.1. Modelling
Due to a simplification in the kinematic model of the Hip joint described in [9]
and multiple errors in reproducing the equations, a more accurate kinematic
model was derived by the first author for the Hip joint. Kinematic equations
were also derived that describe the torque around I due to the spring extension.
The force of the spring was modelled as:
664

Force = (Lr - k ) k + FI
LJ = final length of spring
L = initial length of spring
k = spring stiffness
FI= minimum force to extend spring

2.2. Method of Determining Optimal Parameters

The new kinematic models for the Hip joint and spring were coded in Matlab.
Design characteristics of the springs were matched to available springs that were
thought to have adequate stiffness and be of suitable length for the dimensions of
the leg. Preliminary findings resulted in a spring with twice the stiffness of the
original being chosen. Final specs for the springs were L, = 90 111111, k = 6.54
N/m and F, = 58.86 N.
Several experiments were conducted to determine the available lifting and
downward torque of an unaltered Hip joint. Figure 2 shows the original lifting
torque of the Hip joint at system air-pressure of 6 Bar.
Spring Torque Design for Robug 4 Hip Joint
at 6 Bar Air-Pressure

_
8 ................. ;”........................ :.........: .........:.........:.........
’Y : . . . .
. . ; ....... . : . .
............... ,: ......... ............................
~

-{8 -06 -0.4 -0.2 0 02 0.4 0.6 0.8 1

Theta (Rad)
Figure 2. Spring and Plate Design

The Design Torque was then calculated as the Spring Torque plus a constant
Buffer to ensure adequate lifting torque remained. The Spring Torque was
modified by changing the spring support position, C, until the Design Torque
was as close as possible to the original lifting torque without exceeding it
anywhere along the range of 0. This ensured the maximum possible Spring
 665

Torque without reducing the resultant lifting torque below the Buffer (in this
case the Buffer was set equal to 4 Nm).
Once the spring support position, C, was determined, a mild steel plate was
made that replaced the original link BD and provided support for a mild steel
shaft at C that the springs would be attached to (see Figure 1).

3. Results
The original and increased supporting torque can be observed in Figure 3. The
increase in supporting torque due to the springs is substantial. At 8 = 0 there is
at least 36% more torque available to the Hip. From 8 = 0.2 there is at least 40%
more torque. Since the leg design is of the insect type, where the Knee is usually
located above the Hip joint, the increase of supporting Hip torque for the useable
range of 9 >_ 0, is at least 36%.

Figure 3. Increase in Available Supporting Torque of Robug Iv Hip Joint at 6 Bar Air-Pressure

The increase in supporting torque is irrelevant unless adequate lifting torque

remains to relocate a leg. Figure 4 shows the original and decreased lifting
torque after adding springs. Lifting torque is drastically reduced, but remains
above zero. This ensures the leg can still be lifted, although with an extended
response time. In the preferred operating range of 9 >_ 0 the minimum lifting
torque is 4 Nm. Please note that the torque plotted in Figure 4 remains after the
weight of the leg is accounted for.
666

Theta (Rad)
Figure 4. Available Lifting Torque of Robug IV Hip Joint at 6 Bar Air-Pressure

The designed and actual spring torque can be seen in Figure 5. The two are
remarkably close, and promote confidence in the accuracy of the kinematic and
computer models.
The weight of the new plates springs and shafts total 1.155 kg.
Manufacturing the plates from aluminium will reduce the weight to 0.627 kg or
less. Using hollow stainless steel shafts instead of solid mild steel would reduce
weight further.

4. Observations
The final spring and aluminium plate arrangement would add an extra 3.8 kg at
the most to the 22.2 kg robot. This does not impact much on leg performance, as
the extra weight is near the body; however it does mitigate the 36% increase in
supporting torque (minimum) by a 17% increase in weight. This trade-off is
expected to become more favourable as higher air-pressure is used. This is
because only a small increase in the weight of the springs should be necessary
over the current springs of lower stiffness.
 667

Comparison of Designed and Actual Spring Torque

4 ..+..*.. I ........ j. .... : . ... I .. . . . . I. . ... . I . ... ..;... . .I .....

. .. .
. .,. ..
. . . .

Theta (Rad)
Figure 5. Comparison of Designed and Actual Spring Torque

Increased friction due to the springs can be observed in some plots. Actual
Spring Torque was close to but less than the Designed Spring Torque (see Figure
5). However, Figure 4 shows that the Lifting Torque with Springs is less than
the Buffer of 4 Nm for 8 c 0. This disparity can be explained by increased
friction due to the large tangential force of the spring at pivots D and B . At
most, this friction accounts for a loss of 2 Nm of torque. At higher system air-
pressures this friction will increase, and may require a change of strategy from
nylon bushes to bearings.
While the data for actual spring torque in Figure 5 does follow the shape of
the designed spring torque, it does seem to be affected by a small random factor.
This is most pronounced for 0 > 0.4 and reaches a maximum value of 2 Nm at 8
= 0.5697. This phenomenon can be explained by stiction that is introduced by
the friction between the springs and the shafts they are attached to. This can
cause the springs to bend rather than stretch thus producing a torque opposing
the motion of the system. This could be reduced by separating the spring and
shaft with a nylon bush.

5. Conclusions
Redirecting some of the excessive lifting torque available to a pneumatically
powered leg to assist in supporting the robot is feasible and very beneficial. An
increase in supporting torque of 36% to well above 40% was found for the
Robug IV Hip configuration.
668

Friction and stiction mitigation methods need to be investigated. These

contributors will become more significant at higher system air-pressures. The
importance of this discovery emphasises the need for implementation and
experiment, not just theoretical analysis.
The performance of this system will improve with higher system air-
pressures, provided the friction and stiction factors are reduced. Not only should
higher air-pressure result in a more significant increase of supporting torque than
36%, but they should also experience decreased time for relocating a leg (since
the lifting response time will be improved).
Increasing the power to weight ratio of pneumatic cylinders will
significantly improve the performance of legged robots that incorporate them.
Payload can be increased by the same percentage as the supporting torque
increase, or extra equipment such as a pressure vessel and regulator could free
pneumatically powered robots from their tether, making them much more
mobile. Different gaits could also be possible along with corresponding
increases in speed and stability.

6. Acknowledgment
The authors would like to extend their gratitude to the University of Southern
Queensland (USQ) for funding this research.

7. References:
1. Zielinska, T., Synthesis of Control System - Gait Implementation
Problems. 4th International Conference on Climbing and Walking
Robots, 2001: p. 489.
2. Pratt, J., B. Krupp, and C. Morse, Series Elastic Actuators for High
Fidelity Force Control. Industrial Robot: An International Journal,
2002.29(3): p. 234-241.
3. Colbrunn, R., G. Nelson, and R. Quinn. Modeling of Braided
Pneumatic Actuators for Robotic Control. in 2001 IEEE/ RSJ
International Conference on Intelligent Robots and Systems. 2001.
4. McLatchey, G.J. and J. Billingsley. Force and Position Control Using
Pneumatic Cylinders. in 9th International Conference on Climbing and
Walking Robots. 2006. Belgium.
5. Jiang, W.Y., A.M. Liu, and D. Howard. Foot-force Distribution in
Legged Robots. in 4th International conference on Climbing and
Walking Robots. 2001.
6. Kar, D.C., Design of Statically Stable Walking Robot: A Review.
Journal of Robotic Systems, 2003. 20(11): p. 671-686.
 669

7. Cubero, S., Force, compliance and position control for a pneumatic

quadruped robot, in Faculty of Engineering. 1997, University of
Southern Queensland: Toowoomba.
8. Kingsley, D.A., A cockroach inspired robot with artificial muscles, in
Department of Mechanical and Aerospace Engineering. 2005, Case
Western Reserve University. p. 231.
9. Galt, S., Soft Computing Based Motion Control for an Eight Legged
Robot", Thesis, Portsmouth: , Dept, of, 1998., in Electrical and
Electronic Engineering. 1998, University of Portsmouth: Portsmouth.
 INTEGRATED INTELLIGENT MECHROBOT SYSTEM

LIQIONG TANG AND GURVINDER SINGH VIRK

Institute of Technology and Engineering

Massey University, New Zealand

This paper presents the design and implementation of an Integrated Intelligent

MechRobot System (IIMRS). The framework and the functionality of IIMRS are
developed based on intelligent control, robotics, shop floor integration, and factory
automation. Modern control software and hardware can be easily implemented with
IIMRS. The system consists of autonomous robot, embedded system, PC-based control,
image processing, and functionalities commonly used for robotic control and system
automation. IIMRS aims to develop a real-time physical training and learning
environment for the design, simulation, and testing of mechatronic systems, robotic
control, and industrial automation.

1. Introduction

With the increased computing power and the advances in artificial intelligence,
smart robots are now marching into every possible comer of the industrial field
and our daily lives. Brainless machines and hard automation systems are being
replaced by intelligent systems with sophisticated sensors and flexible
hnctionalities. The design and development of such modem systems sits on the
foundation of mechatronics, robotics, computer systems, artificial intelligence,
machine vision, automation control, etc. An intelligent system, a robot, or a
control system follows the same product development cycle just like any other
product on the market. Design analysis, modifications, testing and retesting are
unavoidable for any new product development. The aim of the Integrated
Intelligent MechRobot System (IIMRS) is to provide a physical platform to
support the design and development of intelligent machines and automation
systems, which is becoming an increasingly complex and difficult task.
IIMRS is an integrated system. It mimics an automated shop floor [l]. The
system consists of two robots, one microcontroller-based autonomous robot, the
other is a 4-axis pick-and-place robot controlled through a PC. A pneumatic
transport frame with two pallets and two robotic grippers is for transferring
components between different locations. The pneumatic system is controlled
using two industrial PLCs. Quality data acquisition and control relies on a
computer-based vision system. All the subsystems in IIMRS can also be linked
through a network. The schematic diagram of the whole IIMRS system is
illustrated in Figure 1. The first version of IIMRS has been successfully
implemented in the Institute of Technology and Engineering at Massey

670
 671

University. It is currently used for undergraduate and postgraduate teaching and

research projects. IIMRS provides a physical environment with features and
functionalities commonly used in robot control, mechatronic system design, and
industrial automation. IIMRS provides a platform a robot, a control system, or
an assembly line to development new control strategies and test new
hnctionalities. It also simulates and motivates designers to further improve
existing systems and develop new ideas. The system has been used in studying
mechatronics applications and the design and testing of autonomous robots.

Gripper
C B

Pallet 1 1 / /’ + J Pallet 1

1’ A Pattern and
( Obstacle ) QC check

v \
\
\
\

n A

n F”, Gripper2
4-axis Robot

Figure 1. The schematic diagram of IIMRS

2. Autonomous Robot Design

Automated Guided Vehicles (AGV) have become common facilities on the

factory floor for transportation and material handling. The autonomous robot
672

running under IIMRS mimics the common functionalities of a real AGV [2], [3].
The “brain” of the vehicle is a microcontroller with embedded programs. It is
designed to be able to perform the functions commonly implemented in an AGV
such as [4],[5]:

1) Auto-recognize part loading and unloading

2) Give warning signals
3) Avoid obstacles on its path
4) Transfer parts to different destinations
5) Accurately position the parts for the robot to pick up

The starting position of IIMRS is located at position A as shown in Fig 1.

At the beginning the autonomous robot will stay at position A waiting for the
part to be loaded. The 4-axis pick-and-place robot is able to sense the
appearance of the autonomous robot. Once the part is placed on top of the
autonomous robot, it will trigger the robot to move from location A towards G
via F. The route from A to G simulates an AGV on an open factory floor
delivering parts to different locations with the possibility of having to overcome
obstacles on the way. Different sensors and methodologies such as proximity
sensors, laser guidance, electrical compass, and beacons are all successfully
used for guiding the autonomous robot to avoid the obstacle and correctly guide
the robot to arrive at position G. The autonomous robot is required to be parked
at position G between the two dark lines for a few seconds. This imitates the
situation of a part under some service. Once the time has elapsed, the robot will
move towards position E by following a defined line to send a part for the robot
gripper at location E to pick up. Such a functionality of an autonomous robot is
often seen on production and assembly lines. The autonomous robot is required
to be able to accurately position the part under the pick-and-place robot at
location E for the gripper to pick up the part. When the part is unloaded, the
autonomous robot will follow the path back to position A and position itself for
the next cycle.
The brain of the autonomous robot is a C8051F020 microcontroller from
Silicon Laboratories (SiLab). The SiLab C8051F020 is a fully integrated
microcontroller with mixed-signal system on a chip. The main features of the
microcontroller are listed in Table 1.
 673

Peak throughput 25 MIPS

FLASH Program Memory 64K
On-chip Data RAM 4352 bytes
Full-duplex UARTS x2
16-bit Timers x5
Digital I/O Ports 64 pins
12-bit lOOksps ADC S channels
8-bit 500ksps ADC 8 channels
DAC Resolution 12 bits
DAC Outputs x2
Analog Comparators x2
Interrupts 2 Levels
PCA (ProgrammableCounter Arrays)

The following example is just one of the many autonomous robots running
under IIMRS. This robot uses proximity detection for obstacle avoidance.
Modulated IR sensors are used as the proximity detection units. The autonomous
robot has two such units, one on each front comer. It is programmed to turn
away from the object when an object comes within 6 cm. If both the proximity
units are activated at the same time, the robot will back away from the obstacle
and turn. The guidance for the robot to move from location A to G via F is
provided by beacons using IR sensors. Each beacon unit has two outputs, one
for directing the robot and the other for stopping the robot when it reaches the
destination. The robot identifies the beacon units through different frequencies.
The receivers on the robot are linked to a microcontroller and the
microcontroller is able to identify which beacon it is communicating with.
Using beacon units make the guidance for the autonomous robot very flexible.
As the beacon units can be placed at different locations, it is easy to expand the
system to have more destinations at different locations for the robot to move
around. In Figure 2, location F and G each has a beacon unit to communicate
with the autonomous robot. The ability of the autonomous robot communicating
to the beacons and then using the beacon’s positions correctly to direct the robot
to location G via F can be easily demonstrated.
674

Figure 2. Communication using beacon units

3. Pick and place robot for loading and unloading

A PC-based pick-and-place robot shown in Figure 3 is integrated in IIMRS for

part loading and unloading [6],[7].A motion control card from National
Instruments is implemented in the robot control system. The motion control
card is a combination of a servo motor and stepper control card. It can provide
full motion control for independent as well as coordinated four-axis motions.
The servo axes always operate in closed loop control. All the stepper axes
support half and full step as well as microstepping possibilities. The driver
software called NI Motion is also from National Instruments, which allows
creating motion control applications using the graphical programming Motion
Assistant. With Motion Assistant, it is possible to transfer the application
program to either a LabView program or a program in other computer languages
such as Visual Basic, C, or C#.
The robot has 4 motors and the motion control card can cope up to 4-axis
control. The 4 motion control axes for the robot are defined as:

Motion control axis 1 - X axis (robot base rotation)

Motion control axis 2 - Y axis (lower a m rotation)
Motion control axis 3 - 2, axis (upper arm rotation)
Motion control axis 4 - Gripper movement
 675

Figure 3. A 4-axis PGbased pick-and-placerobot

A software program for the control of the robot is developed using LabView.
The control panel for the robot arm is presented in Figure 4. The control system
has two modes, automatic and manual, which can be toggled through the push
button on the panel. The control panel also provides a user friendly interface to
teach the robot and define the pick and place positions. There are two indicators
on the panel, one for indicating that a part has arrived at location A and another
for signalling the autonomous robot is ready for loading. Only until both
indicators are switched on can the robot arm be activated. In other words, the
Boolean “AND” of the two indicators trigger the pick and place movement of
the robot arm.
676

Figure 4. Pick and Place Robot Control Panel

The robot works closely with a part data collection system. Based on the data
of the part such as geometry and dimensions, the robot is able to identify
differences and perform part sorting functions. Such processes are commonly
used for product quality control. In IIMRS, each time a part is delivered to
location A by pallet I, the robot will load the good part on top of the
autonomous robot and put the bad ones into a bin.

4. Part Handling System

The part handling system is a pneumatic system controlled by two Omron

Programmable Logical Controllers (PLC). The pneumatic frame consists of two
pallets and two pneumatic grippers refer to Figure 1. Apart from picking and
placing functions, one of the grippers can also rotate to place components at
different orientations. Each PLC controls one pallet and one pneumatic gripper.
Sequential control is implemented in the control system. The pneumatic system
is activated by the autonomous robot. Each time the autonomous robot amves
at location E and correctly positions the component under &per 2, the
pneumatic system will be trigged by a sensor. Griper 2 will come down and
pick up the part and travel to Location D to place it into the holder on top of
Pallet 2. After Gripper 2 is returned, Pallet 2 will move the part from Location
D to C. The controls are realized through an Omron PLC. The corresponding
 677

state diagram is illustrated in Figure 5 . The other PLC is for the actions from
Location C to A via B. The process functions are very similar except the griper
is able to rotate up to 90 degrees. Once the part is transfer to Location C,
Gripper 1 will pick up the part and rotate 90 degrees first, and then transfer the
part to Location B. When Gripper 1 puts the part in Holder 1, the part is in a
new orientation. Pallet 1 will then send the part to Location A.

Figure 5 . The state diagram for the control from position E to C

5. Real-time data collection

A real-time data collection system is implemented in IIMRS. The system is

running on a PC. Each time when a part is transferred from Location B to A, it
will pass the data collection point where the information of the part is captured
through different sensors or by a camera. The part data is processed in real time
and then output the control information the pick and place robot at Location A.
If the part does not have the correct geometry or wrong size, the robot at
Location A will pick it out. Otherwise it will be put on top of the autonomous
robot to activate the next cycle.
678

The following example uses a camera in the data collection system. While
pallet 1 travels from position B to A, the system controlled by the PC will
recognize the badly made components. When a problem part is recognized, the
system will give an alarm signal. To implement such a system a computer-
based vision system is developed. This system visually checks whether parts on
pallet 1 moving from position B to A are faulty or not. This vision system uses
a Logitech QuickCam webcam with a USB plug. The system is developed
under Visual Basic 6.0 (VB).When the VB program is run the system creates a
capture window to display the streaming video from the webcam. Through
Microsoft WDM Image Capture driver it is possible to get the image from the
webcam to the capture window. Once the image has been taken of the part the
image processing is immediately done to extract the data from the image. Part
length, width, and area are the three variables to be measured. The whole image
is scanned and any pixel less than a preset variable is the part and coloured red
and any other pixel is said to be background and set to light blue.
A Graphical User Interface (GUI) is designed to give a visual display of
what is happening. Figure 7 shows a screen capture of the program running and
a correct part passed through,

Figure 7: Screen capture of a good part. Figure 8: Screen of a part that was short.

The top left window shows the live streaming image from the webcam, the
top right window shows the captured image that is to be processed. The bottom
left window shows the processed image and note the shape of the bolt has been
brought out in red, The bottom right frame show data about parts being passed
through. It displays the number of parts passed through, showing parts failed and
parts passed. Also for each image that is processed it displays the value
calculated for bolt length, bolt width and bolt area. Then there is a large clearly
 679

visible text area which displays the result of each part processed as shown in
Figure 7 and 8.

6. Conclusions

The framework and the functionality of IIMRS has been developed and
implemented in the Mechactronics Labortory at the Institute of Technology and
Engineering, Massey University. The system covers the use of microcontroller,
PLC, pneumatic equipment, motor control, sensing system, vision, and
intelligent control. It provides a useful platform for engineering students to
study both hardware and software. It is also an ideal test rig for robot design,
automated pneumatic system, simulation of shop floor control, and intelligent
factory automation. The system is now used for Mechatronics and engineering
undergraduate teaching and as a test rig for autonomous robot design projects.

Acknowledgments
The authors are grateful to all the 4” year mechatronics students who
contributed to IIMRS system. In particular, thanks are due to: John West, James
Barkwith, Lyndsy Homes, Lara Christian, Glenn Gregory, Alex Wishart, Kieran
Orchard, Rob Paddison.

References
1. J. A. Rehg, “Introduction to robotics in CIM systems”, Prentice Hall (2000).
2. Robert J. Miller, “Robotics : future factories, future workers”, Sage
Publications (1983).
3. Joseph A. Angelo, “Robotics: a reference guide to the new technology”,
Greenwood Press (2006).
4 . T. Braunl, “Embedded robotics : mobile robot design and applications with
embedded systems”, Springer (2003).
5 . U. Nehmzow, “Mobile robotics : a practical introduction”, Springer (2003).
6. John J. Craig, “Introduction to robotics: mechanics & control”, Addison-
Wesley Pub. Co. (1986).
7 . 7340 User Manual, National Instruments (2003).
MCA2 - AN EXTENSIBLE MODULAR FRAMEWORK FOR
ROBOT CONTROL APPLICATIONS

Klaus Uhl' and Marco Ziegenmeyer

FZI Forschungszentrum Infonatik, Interactive Diagnosis and Service Systems (IDS),
Haid-und-Neu-Str. 10-14, 0-76131 Karlsruhe, Germany
*E-mail: uhl@fzi.de

For the rapid development of different robot prototypes the adoption of

reusable software components with standardized interfaces is a very impor-
tant aspect. In this context reusability can mean reuse on a different robot or
in a different environment (e.g. user space vs. realtime). MCA2 provides an
extensible modular framework for robot control applications in order to fulfill
these needs.

Keywords: Control architecture; Robots

1. Introduction
Efforts in software development should be restricted to create new compo-
nents and - in the context of robotics - new control methods which have
not been implemented so far. A framework which encourages the devel-
oper to organize the whole controller system as a set of small modules with
standardized interfaces which can be individually reused in other projects
or other execution environments is a huge step into this direction. If the
communication between the modules and the synchronization between the
different parts of the software are implemented by the framework then the
developers can focus their work solely on the methods and algorithms which
are necessary for controlling the robot.
The Modular Controller Architecture Version 2 (MCA2)l was devel-
oped to meet these requirements. It is based on MCA2 but provides addi-
tional enhancements. MCA2 is written in C++, runs on Linux and Windows
and uses RTAI/LXRT3 for hard realtime operation. The graphical tools are
written using Qt.4
This paper presents the architecture and the core components of MCA2,
introduces its tools and shows some projects which were implemented using

680
 68 1

MCA2. In addition it gives an overview of other frameworks for robotic

applications.

2. Architecture
MCA2 targets at the design and development of low and medium level robot
control software. Therefore it uses a strictly hierarchical control structure
and a two-way data flow. The lowest level comprises the hardware interface
while the units get more and more abstract when you ascend the system
hierarchy. The data flow is split into two parts:

Sensor data is flowing upwards from the hardware interface units

through all units which need to process it. Intermediate units may
alter the sensor data or create new "virtual" sensor data from it.
Control data is flowing downwards from the highest level control
units or the user interface bwk t o the hardware interface units.
Control data may also be altered or derived by intermediate mod-
ules.

Each unit can participate in both the sensor and the control data flow.
The basic units of an MCA2 application are Modules, Groups and Edges.
Modules are the smallest units and contain basic functional blocks. As
complex control systems usually contain many functional blocks which are
combined t o even bigger entities modules can be put together into groups.
They can be connected with edges and are then able t o communicate with
each other.

2.1. Modules
Figure 1 shows the schematic structure of a module. Sensor data flows into a
module through the Sensor Input interface. It is processed in the Sense0
function which writes its results int,o the Sensor Output interface. From
there the sensor data flows upwards to the higher levels of the system hi-
erarchy. On the other side the control data flow enters a module through
the Controller Input interface. Control data is processed in the Control()
function. It writes the altered or newly created control data to the Con-
troller Output interface from where the control data flows further down the
system hierarchy.
The Sense 0 and Control 0 functions can exchange information via
local variables which comprise the internal state of the module. They are
682

called periodically by the control loop of the enclosing thread container (see
section 2.4) in configurable time intervals.
The four sensor and controller interfaces are modeled as vectors of type
double. Each interface entry has a description which defines its meaning.
Furthermore each interface owns a changed flag so that Control0 and
Sense0 can determine in each cycle of the control loop whether the con-
troller or sensor input needs to be processed.
As many algorithms depend on certain parameters which are mostly
static and are not influenced by other modules each module has a Parameter
interface. In contrast to the sensor and controller interfaces parameters are
not restricted to double values but can be boolean values, integer values,
floating point values or strings. Parameters can be set from within the
MCA2 application itself, e.g. with values read from a configuration file.
Another more powerful option to alter parameters is the browsing and
debugging tool described in section 3.2. With this tool it is possible to
access and change parameters at runtime and thus fine-tune the application
without the need to restart it.

Fig. 1. A Module with its five interfaces (left) and a Group containing several commu-
nicating modules in a two-layer hierarchy (right).

2.2. Groups
To build a bigger functional block from basic functional blocks which are
implemented in individual modules an instance of each module is put into a
group. Then the modules are connected by Edges. An edge always connects
an output interface of one module with a corresponding input interface of
another module (controller output with controller input and sensor out-
put with sensor input). Either complete interfaces or parts of them can be
connected. Arbitrary permutations of interface elements are allowed. Edges
carrying sensor data are always directed upwards while edges carrying con-
trol data are directed downwards as can be seen in figure 1.
 683

The graph of modules and edges within a group can be organized in a

multilayered hierarchy in the following way:

0 All modules which either have no connected sensor input and con-
troller output interfaces or which are only connected t o the group’s
sensor input and controller output are placed on the lowest level
1 := 0.
0 Each module is placed one level above the highest level module t o
which its sensor input or controller output is connected.
0 Modules on the same layer are arranged in the order in which they
were added to the group.

Groups behave like modules implementing bigger functional blocks.

Therefore they also have controller input, controller output, sensor in-
put and sensor output interfaces (the parameter interface is not used for
groups) and can themselves be nested into other groups t o build even bigger
functional blocks. Moreover, groups also have a Sense0 and a Control0
function. They manage the processing and data flow inside the group. The
Sense 0 function performs the following steps:

(1) Copy the group’s sensor input data to all modules connected t o it by
edges.
(2) Set the current level 1 := 0.
(3) Call the Sense0 function in all modules on the l’th level.
(4) Copy the sensor output data of all modules on the l’th level which
have set their sensor output’s changed flag to the modules which are
connected t o them by edges. If there is an edge t o the group’s sensor
output then copy the data there.
+
(5) Set 1 := 1 1.
(6) If there are modules on the l’th level go to step 3. Otherwise processing
is finished.

The group’s Control0 function performs virtually the same steps but
starts with the group’s controller input data and the highest level mod-
ules. It then descends the module hierarchy and passes data from controller
output t o connected controller input interfaces.

2.3. Parts
A Part is the basic execution environment of an MCA2 application. It either
manages a single module or - usually - a group. It initializes the execution
684

environment, communicates with other parts over a TCP/IP connection

and runs the control loop of the main thread.
Parts expose the interfaces of the managed group to other parts which
can use this information to establish inter-part edges. These are similar to
regular edges as described in section 2.2 but connect sensor (or controller)
outputs of one part with sensor (or controller) inputs of another part.
A part’s control loop is executed periodically in configurable time inter-
vals and consists of the following steps:
(1) Pass all sensor data received from lower-level parts to the sensor input
interface of the managed group.
(2) Call the Sense0 function of the managed group.
(3) Pass the sensor output data of the managed group to all connected
higher-level parts.
(4) Pass the control data received from higher-level parts to the controller
input interface of the managed group.
(5) Call the Control() function of the managed group.
(6) Pass the controller output data of the managed group to all connected
lower-level parts.

2.4. Multithreading
Thread containers allow for splitting an MCA2 application into multiple
concurrent threads of execution. Like parts thread containers manage a
module or a group and expose their sensor and controller interfaces to
other modules. Additionally they run their own control loop in a separate
thread. The control loop is similar to the part’s control loop. But instead of
sending and receiving data from other parts the interfaces of the managed
module are only copied to the exposed interfaces of the thread container so
that a thread-safe access is guaranteed.

2.5. Realtime and network transparency

MCA2 offers complete realtime and network transparency. From the view-
point of the implementation of groups and modules it makes no difference
where a module is executed, be it in the same thread, in the same process,
in a different process or even on a different host. It can also be run either
in a realtime or a non-realtime thread without modifications.
This transparency makes it possible to develop a controller system in a
standard Linux or Windows environment using the usual set of development
tools like compilers, editors and debuggers. Most bugs can be fixed with
 685

those tools before the system is run in a realtime environment. Due to the
use of RTAI with its LXRT extension as the realtime operating system
of choice it is only a matter of setting special thread priorities to execute
specific thread containers in hard realtime.

2.6. Blackboards
The controller and sensor interfaces are modeled as vectors of double values
and are therefore only suited for small to mid-size data. But sometimes there
is the need to exchange big and complex chunks of data between modules
which may even dynamically change its size. MCA2 contains blackboards
for this purpose.
A blackboard is merely a memory buffer which can be accessed as a
vector of elements with a user specified size. The blackboard elements are
not typed but can be converted into anything the application demands. The
only restriction is that the types which should be stored in blackboards must
be placable in a continuous memory block.
Blackboards are also network transparent and can be used in any part
of the system whether they were originally created in the same process or
on a different host. The access to blackboards is protected with read-write
locks which work even across host boundaries.

3. TO013

A framework is only as good as the tools which it provides to assist devel-

opers. Because of this MCA2 comes with two powerful tools for browsing,
debugging and controlling MCA2 systems. Both tools connect to an MCA2
system via TCP/IP and can therefore also be used from a distant location
for remote operation or diagnostics.

3.1. User interface

The mcagui allows to quickly click together a control user interface and
to connect it to an MCA2 system. The mcagui runs atop the highest-level
part in the system, reads its sensor outputs, writes its controller inputs and
accesses its blackboards. It comes with a set of general-purpose widgets like
buttons and value displays but can also be extended through plugins which
provide special functionality for a certain application.
686

Fig. 2. The rncagui showing the control interface for the mobile research platform Odete.

3.2. 3rowsing and d ~ b u g g ~ n g

The ~ c a b r o w s ~isra browsing and debugging tool. It can be used to browse
the structure and components of a running MCA2 system. This includes all
modules, groups, edges, parts, component interfaces and blackboards. One
can inspect the current values of interface entries and alter them even while
the system is running. Additionally individual modules can be stopped
and restarted, debugging messages which are written to the console can be
switched on and of on a per-module basis and the current debug level, i.e.
the amount of debug messages that are generated, can be changed.

Fig. 3. The mcabrowser showing the module graph of a group, a module with its in-
terfaces, parameters and blackboards and an interface window to view and alter the
controller output of a module.

4. Projects
During the years MCA2 has been successfully used for many robotic
projects within our group and in other research groups. Some of those
 687

projects have made their way into industrial applications which are de-
ployed worldwide. Figure 4 shows a collection of current robotic projects
which have been developed using MCA2.

(a) The autonomous sewer inspec- (b) The six-legged walking machine
tion robot MAKROplus.'

..
"Odete" with a "Puma 200" manip- matic guided vehicle "LTC2" from
ulator. Swisslog Telelift.

(e) The autonomous "Smart Road-

ster" .? perception-action integration Ar-
mar m.*
Fig. 4. Robotic projects which were implemented using MCA2.

5. Related work
Besides MCA2 there exist several other frameworks for robotic applications.
Each of them takes a slightly different approach and sets a different focus.
In the following a collection of popular frameworks is presented. In contrast
to MCA2 they all don't support hard realtime operating systems and are
thus not suited for robot control systems in which at least certain parts
688

need to run in a predictable realtime environment.

ORCA' follows a Component Based Software Engineering (CBSE) ap-
proach to develop robotic systems. ORCA provides the developer with a
framework for developing components which communicate via interfaces.
ORCA does not define a system architecture but only provides a means t o
build components and implements the component communication. ORCA
comes with a library of standard components but has no powerful browsing
and control tools.
In Mzro" components are implemented as CORBA services. On the
lowest level Miro is less more than an easy to use wrapper around CORBA.
But Miro adds several layers atop the basic system which provide generic
services for communication, sensor and actuator interfacing and so-called
"frameworks" which can be specialized or extended for given robotic tasks.
Although Miro comes with several graphical development tools none of
them is able to inspect or modify a running system.
The Player/Stage/Gazeboll project consists of the network transparent
robot device server Player, the lightweight 2D robot simulator Stage and
the high-fidelity 3D multiple robot simulator Gazebo. It is designed to sup-
port virtually any architecture and language for developing robot control
applications. But Player's support ends a t the device layer. It provides no
support for higher-level components.
MARIE12 also follows a CBSE approach but unlike most other robotic
frameworks it mainly acts as a mediator between components from differ-
ent frameworks. The advantage of this approach is that the best available
implementations of the needed functions can be tied together to build a
robot control application. The drawback is the large communication and
computation overhead and the big memory footprint.

6 . Summary and outlook

In this paper MCA2, a scalable modular controller architecture for robotic
applications, has been presented. It allows to easily reuse components and
thus makes the process of getting new robot prototypes working faster. De-
velopment and testing can be performed in Linux or Windows user space
before the system is deployed to a realtime environment. The network trans-
parent communication between individual parts of a complex controller sys-
tem makes it possible to distribute such a system over several hosts without
any additional effort. The user interface mcagui provides a fast and flex-
ible way to control a system and visualize data from it. The use of the
mcabrowser enables the developer to browse the complete system while it
 689

is running and fine tune system parameters.

Although MCA2 has been successfully used for many robotic projects
some aspects of the design of MCA2 can still be improved. The next major
release will include typed interfaces and blackboards. Moreover an event
mechanism for asynchronous communication will be introduced.
MCA2 is released under the terms of the GNU General Public License
and is available at http://www.rnca%.org.

References
1. Modular controller architecture version 2 http://www.mca2.org.
2. K.-U. Scholl, J. Albiez and B. Gassmann, MCA - An Expandable Modular
Controller Architecture, in 3rd Real- Time Linux Workshop, 2001.
3. RTAI - the RealTime Application Interface for Linux http://www.rtai.org.
4. Qt - The Cross-Platform C++ Development Framework
http://www.trolltech.com/products/qt.
5. C. Birkenhofer, S. Studer, J.-M. Zollner and R. Dillmann, Hybrid impedance
control for multi-segmented inspection robot kairo-ii, in International Con-
ference on Informatics in Control, Automation and Robotics (ICINCO 2006),
2006.
6. B. GaBmann, Modellbasierte, sensorgestutzte navigation von laufmaschi-
nen im gelande, PhD thesis, Fakultat fur Informatik, Universitat Karlsruhe
(TH)2007.
7. J. Schroder, U. Muller and R. Dillmann, Smart Roadster Project: Setting up
Drive-by-Wire or How to Remote-Control your Car, in Proceedings of the 9th
International Conference on Intelligent Autonomous Systems, 2006.
8. T. Asfour, K. Regenstein, P. Azad, J . Schrijder and R. Dillmann, ARMAR-
111: A HUMANOID PLATFORM FOR PERCEPTION-ACTION INTE-
GRATION, in HCRS, Second international workshop o n Human-Centred
Robotic Systems 2006, October 2006.
9. A. Makarenko, A. Brooks and T. Kaupp, Orca: Components for robotics, in
2006 IEEE/RSJ International Conference on Intelligent Robots and Systems
(IROS’OS), Dec 2006.
10. H. Utz, S. Sablatnog, S. Enderle and G. Kraetzschmar, IEEE Transactions
on Robotics and Automation, Special Issue on Object-Oriented Distributed
Control Architectures 18,493(August 2002).
11. B. Gerkey, R. T. Vaughan and A. Howard, The player/stage project: Tools
for multi-robot and distributed sensor systems, in Proceedings of the 11th
International Conference on Advanced Robotics ( I C A R 2003), June 2003.
12. C. CBtB, Y . Brosseau, D. Lktourneau, C. Ra’ievsky and F. Michaud, Interna-
tional Journal of Advanced Robotic Systems 3,55(March 2006).
 MOTION ESTIMATION AND SELF-LOCALIZATION BASED
ON COMPUTER VISION AND ARTIFICIAL MARKER
DEPOSITION

SAVAN CHHANIYARA, KASPAR ALTHOEFER, LAKMAL D SENEVIRATNE

Department of Mechanical Engineering, King's College London
Strand, London WC2R 2LS, UK

This paper presents a new approach of self-localization utilizing artificial markers

deposited from the mobile robot during motion. The main idea of this system is to
temporarily place artificial markers in the environment such as the surface the robot is
maneuvering over. Once the markers are placed in the environment, they have zero speed,
i.e. they represent the speed of the environment over which the robot is traversing. The
speed of these markers with respect to the moving robot is then measured using a two
dimensional (possibly three dimensional) sensor affixed to the robot. The proposed sensor
system is composed of two main subsystems. The first subsystem generates the markers
and places them on the surface. The second subsystem is a receiving element, which
continuously acquires relative position signals from the markers placed on the surface.
This new sensor concept is envisaged to be applied in the following areas, automotive
sector, planetary exploration and underwater seabed exploration. This approach also can
be extended easily for walking machines. Initial experiments employing a camera system
as sensor have been conducted and results are presented.

1. Introduction
To accurately estimate the motion employing on-board sensors is still a very
important research topic for both mobile & walking robot applications. At
research centers world-wide, efforts aim at developing alternatives to
conventional sensors i.e. wheel odometer, GPSDGPS and inertial sensors. Over
the past decade, several researchers have made significant advances in utilising
various vision and image processing techniques for navigation, obstacle
avoidance and control of ground vehicles [l]. Most of the recent approaches
employ forward facing cameras and use image processing techniques for
generating 2D and also 3D spaces from 2D images, detecting obstacles or
extracting motion from the image sequence [2]-[7]. Different methods can be
used to calculate the relative motion based on the translation and rotation of
points in these 2D or 3D spaces [8]. Although significant improvements have
been made in motion estimation from 3D points (including optical flow, Stereo
Odometry), it has been noted that a considerable amount of processing time is

690
 691

spent on point matching and outlier removal [9]. In the field of walking
machines, it is very important to reach target location accurately and
autonomously. Problem of localization get worse in walking machines as
slippage and small errors in position & orientation can lead to failures. Several
sensors and techniques been used for navigation and localization, like laser
sensors, inertial sensors and vision [ 161 & [ 181.

2. Recent Relevant Research

In [lo-141, researchers introduced dropping of artificial markers in to the
environment. Mostly, these approaches addressed the issue of robotic
exploration in knownlunknown environments, where no distance or metric
information taken in to consideration. The approach taken in this paper and by
few research groups is to use droppable markers in to the environment and
localize the robot pose [13-141. In [13], main focus was to develop evidence
grid map and topological maps. The main assumptions in [13] for perceived
position of artificial markers were based on robot’s global position using wheel
encoders. These can generate large error in accurate position estimations. In
[14], author presented a system capable of navigation without prior beacon
locations for autonomous underwater vehicle. Global translationhotation errors
were around 2 meters and a few degrees of heading error, which is too large for
terrain based robot localization. In the field of walking machines, popular
humanoid robot ASIMO [17] uses artificial landmarks arranged forehand on
ground for navigation using camera looking down. These arrangements can
constrain robot motion.

Spatlotemporal Image
Acquisition system

Figure 1 Sensor Block Diagram

3. SENSOR CONCEPT

3.1. Sensor architecture

This paper proposes a novel sensor concept to measure a vehiclehobot’s
trajectory and motion. It is envisaged that the sensor system is mounted on the
robot. Sensing comprises a 2-stage process. In a first step, the system ejects
small, but distinct markers which attach themselves to the environment the robot
692

is operating in. In a second step, the on-board sensor which is designed to be

particularly receptive to the strewn markers senses the distance and orientation
of the robot with respect to these markers see Figure 1. This paper focuses on
the research towards such an active sensor system for the installation on a
ground vehicle. Real-time image processing algorithms are developed to locate
the prominent marker features in the successively acquired camera images and
allow the on-line calculation of relative vehicle speed, position and orientation
with rn accuracy. Initial experiments have been conducted and results are
presented.

3.2. A ~ ~ c iMarker
al selection and discharge
Here in this paper, we propose to place specially designed objects ‘Artificial
Markers’ in the working environment and subsequently and used as landmarks.
In an attempt to simpliEy and speed up the process of locating the markers in an
image, circular (rotationally invariant) markers have been chosen in this
implementation. This proves to be more computationally efficient. One issue
that needs consideration is that the environment is actively altered through the
sensing process and depending on the application the appropriate type of marker
and on board marker dropping mechanism needs to be selected. In most cases, a
temporary (decaying) marker is preferable. The marker needs only to be active
during the acquisition process by the on-board sensor. After this it may decay by
itself or be actively removed by the vehicle.

Figure 2 Image of
marker used in this
study (top left),
simplified represe-
ntation of marker image
used as kernel (top
right), input image for
convolution and
convoluted image plot
indicating 4 clear picks
(bottom right).
 693

3.3. Feuture Zden~~cation

and Trucking
The core of the developed program to detect the markers in the image is a 2-D
convolution algorithm [l5]. Since it is known which markers are used for this
process and at what distance from the camera they will appear, the dimension
and size of the marker image can be computed and this knowledge can be
exploited to create the appropriate convolution kernel (filter mask). The markers
used in this study are circular paper shavings as shown in Figure 2.
Circular markers are rotational invariant. Thus, a single convolution kernel
(which is effectively a simplified, graphical representation of a marker image) is
sufficient to give a strong response at a marker’s location, even if an image
rotation has occurred. The convolution process outputs an image whose peaks
represent their locations of the markers in a camera frame. Also, an intensity
value is stored for each marker, which can be used for distinguishing between
markers if different colour markers used in process.

The following standard convolution algorithm has been employed:

C(i, j ) = (m,n)* B(i - m, j - n )

m=O n=O

Markers

Figure 3 (a) Distances between marker points in two subsequent frames. Distances between all
marker point pairs are shown; matching distances between frames are highlighted. (b) Angle
between marker points in one frame and angle between marker points in second frame are
compared and this function is minimized to obtain the vehicle’s change of orientation from one
frame to another.
 694

3.4. Marker Grouping

In a second stage of the proposed process, the markers which are identified in
one frame need to be paired with the corresponding markers in the subsequent
frame.
An algorithm has been devised that attempts to pair all marker points in the
first frame with all marker points in the second frame and picks those pairs
which would lead to a “reasonable” movement. The first feature is the distance
between marker points in the first frame. Only those pairs of marker points in
first frame can be considered which have a corresponding pair of markers in the
second frame which are the same distance apart. The program calculates the
distance between each pair of marker points in the first frame. This set of
distances is then compared to the set of distances computed for the 2nd frame.
Marker points pairs whose distance varies strongly between frames are excluded
from the further process see Figure 3.
To increase the robustness of the approach, the rotations between the
straight lines spanned by the selected pairs of marker points in the first frame
and corresponding straight lines in the second frame are compared to the
computed rotation of the previous frame. The marker pair with the least change
in orientation is selected. Hence the algorithm eliminates marker pairs which
would lead to an unlikely abrupt rotation from previous frames to the current
frame transition.

4. Experiments & Results

The main objective of the experimental procedure is to prove the feasibility of

the proposed approach. A number of tests were conducted in a lab environment
* in order to establish the feasibility and accuracy of the chosen approach. Tests
were performed using a linear test rig as well as on a mobile robot Pioneer P3Dx
for longer distances, Results are summarized in Table 1 & Table 2 below and
trajectory plot of first experimental set are shown in Figure 4 & Figure 5
respectively.
Tablel: Result Summery for Test rig Experiment Data set

d(m) d’(mm) Avg. Encoder Avg.Marker % Error

Encoder Marker % Error velocity v (mm) Velocity v’(mm)
516.21 530.97 2.85 13.489 13.56 0.526
444.65 442.40 0.506 18.8 19.2 2.13
454.75 449.63 1.125 19.50 20.106 3.01
426.23 429.12 0.678 24.56 24.12 1.79
439.47 435.54 0.894 32.65 32.24 1.255
440.94 434.89 1.37 42.22 41.88 0.805
 695

Trajectory Plot

000 500 1000 1500 2000 25W 3000 3500 4000

Time($=)

Figure 4 (a) Test Rig (b) Estimated trajectory and encoder readings on test rig experiment setl

-50; ,
500 1000 1500 2
Distance in X direction (m)
. .
Figure 5 (a) Mobile robot with marker dropping mechanism (b) Estimated trajectory and encoder
readings on mobile robot experiment setl

Table2: Result Summery for Robot Experiment Data set

dm (-1 4 (-1
Marker Distance Robot Distance % Emorme
1775 1776.1 -0.0619
1775.37 1786.1 -0.6008
1862.002 1897 -1.8449
1867.903 1825.367 2.33027
1880.8 1822.9 3.17626
696

4.1. Discussion of results

The position estimates computed by our system are shown in table I and
compared with on board encoder of Linear Test rig. The results from the
experiments show that high accuracy can be achieved at each step. The average
error over the number of experiments is 1.237%. The error in average velocity
of the carriage motion is 1.6% (Table 1). The results from the mobile robot
experiments are also promising. The proposed system correctly computed the
distance with only a relatively small error. The overall error is less then 3.2%.

5. Conclusion
The proposed sensor concept proved to be fairly robust and responsive to the
markers even when experiencing changes in environment, lighting conditions
and image quality. It is observed that even using a standard web camera with a
low frame rate of 5-20 f p s , good results were achieved. These initial
experiments suggest that the proposed artificial marker disposition concept is
feasible and robot position and speed can be accurately and robustly estimated.

6. Future work
Research is underway to improve affine motion estimation between frames and
to develop a robust self-localization system. Emphasis is also given to
minimizing computational time. Further efforts will aim at improving the
tracking accuracy and analyzing the effect of height variation on velocity and
position estimation. Further, the approach will be extended to walking machines
for both indoor as well as outdoor applications.

7. References
1. DeSouza GN, Kak AC, " Vision for mobile robot navigation: A survey
", IEEE Transactions On Pattern Analysis And Machine Intelligence 24 (2):
237-267 Feb 2002
2. A. Davison, "Real-Time Simultaneous Localization and Mapping with a
Single Camera," IEEE International Conference on Computer Vision, pp.
1403-14 10,2003.
3. Y. Takaoka, Y. Kida, S . Kagami, H. Mizoguchi and T. Kanade, "3D Map
Building for a Humanoid Robot by using Visual Odometry," IEEE Int.
Conf. on Systems, Man and Cybernetics, pp. 4444-4449,2004.
4. E. Marchand, P. Bouthemy, F. Chaumette and V. Moreau, "Robust real-
time visual tracking using a 2D-3D model-based approach," IEEE
 697

International Conference on Computer Vision, pp. 262-268, September

1999.
5. B. Jung and G . S. Sukhatme, “Detecting Moving Objects using a Single
Camera on a Mobile Robot in an Outdoor Environment,” In the 8th
Conference on Intelligent Autonomous Systems, pp. 980-987, March 2004.
6. M. Betke, E. Haritaoglu and L. Davis, “Multiple Vehicle Detection and
Tracking in Hard Real Time,” IEEE Symposium on Intelligent Vehicles, pp.
35 1-356, September 1996.
7. J. M. Ferryman, S. J. Maybank and A. D. Worrall, “Visual Surveillance for
Moving Vehicles,” In Journal of Computer Vision, pp. 187-197,2000.
8. Niko Sunderhauf, Thomas Krause, Peter Protzel, “Bringing Robotics Closer
to Students -A Threefold Approach,” Proceedings of the 2006 IEEE Int.
Conf. on Robotics and Automation, Orlando, Florida - May 2006.
9. David Fernandez and Andrew Price, “ Visual Odometry for an Outdoor
Mobile Robot”, Proceedings of the 2004 IEEE Conference on Robotics,
Automation and Mechatronics, Singapore, 1-3 December, 2004
10. Dudek G. ; Jenkin M. ; Millios E. ; Wilkes D., “Robotic exploration as
graph construction”, IEEE transactions on robotics and automation, Vol. 7,
NO. 6, pp. 895-865, 1991
11. S. Caselli, K. L. Doty, R. R. Harrison, F. Zanichelli, “Mobile Robot
Navigation in Enclosed Large-Scale Space”, Proceedings IECon94,
Bologna, Italy, Sept. 5-9, 1994.
12. Ioannis M. Rekleitis, Vida Dujmovic, Gregory Dudek, “Efficient
Topological Exploration.”, IEEE International Conference on Robotics and
Automation, Detroit, Michigan, USA, pp. 678-681, May 1999,
13. Maxim A. Batalin, Gaurav S. Sukhatme, “Efficient exploration without
localization”, IEEE International Conference on Robotics and Automation,
Taipei, Taiwan, September, 2003.
14. B. Tovar, S.M. LaValle, R. Murrieta, “Locally-optimal navigation in
multiply-connected environments without geometric maps”, IEEE/RSJ
International Conference on Intelligent Robots and Systems, (2003).
15. John C. Russ, “The Image Processing Handbook “,Published 2002, CRC
Press, ISBN 08493 1142X
16. S. Thompson, S. Kagami, and K. Nishiwaki, “Localisation for autonomous
humanoid navigation,” in In Proc. IEEE-RAS International Conf.
Humanoid Robots, pp. 13-19,2006.
17. K. Hirai, M. Hirose, Y. Haikawa, and T. Takenaka, “The development of
honda humanoid robot,” in In Proc. IEEE International Conf.Robotics and
Automation, pp. 1321-1326, 1998.
18. K. Sabe and M. Fukuchi and J. S. Gutmann and T. Ohashi and K.
Kawamoto and T. Yoshigahara, “Obstacle avoidance and path planning for
humanoid robots using stereo vision,” in In Proc. IEEE International Conf.
Robotics and Automation, 2004.
 NEW STANDARDS FOR NEW ROBOTS

GURVINDER SINGH VIRKt

CLAWAR Ltd, U K and Massey University
Institute of Technology and Engineering, Wellington, New Zealand
Email: gsvirk@clawar.org, g.s.virk@massey.ac.nz

Robotics is a relatively new discipline, which came into existence when the first
production line robots were commercialised by Unimates in 1962. Since then the
industrial robot market has matured but has been restricted to manufacturing
applications where humans and robots have been kept largely separated due to safety
concerns. As part of this, industrial robot standardization has been developed to support
the sector. In recent years new service robots and new service robot applications have
been emerging; these do not fit into the “so-called” industrial environments and
scenarios and there are no standards for these new situations. This is causing a bottleneck
in the proper development of robotics and the International Standardization Organisation
(ISO) has recently set up new initiatives to address these concerns. This paper focuses on
these latest activities in robot standardization for the emerging areas of service robotics
where close human-robot interactions are essential. In particular, the work of the new
IS0 TCl84/SC2/F’T2 Project team on Robots in personal care and the Advisory Group
AGl on Service robotics are presented.

1. Introduction
It is widely acknowledged that robots will soon begin to play a significant role in
our everyday lives to assist us in performing a huge range of service and
personal care tasks. As such, robotics is seen as likely to become a booming
business area and major companies, such as Microsoft, have already started to
invest in it (see Gates [ l ] in January 2007 issue of Scientific American,
www.sciam.com). UNECE and IFR have forecast that the worldwide market for
robots will be in excess of $66B by 2025, [2]. The main barrier to rapid growth
is that there are no standards (especially no safety standards) that cover the new
service robotic sectors and organisations fear litigation in the event of accidents
occurring (see Virk [3]). In addition there are some ethical issues arising in the
use of robots in the new sectors (see Virk et a1 [4]). Current robot standards only
apply to “industrial robots” but these do not extend to the new types of robots

Work partially supported by the U K s DTI.

698
 699

being developed for the new applications and non-industrial environments

because the design and operational aspects are quite different.
The current IS0 10218-1 safety standard [ 5 ] has been recently revised to
update the safety requirement for industrial robots so that some of the concerns
can be addressed. The changes includes the provision of a “collaborative” robot
operational mode where robots and humans can work together. It states that the
robot must provide a visual indication when it is in the “collaborative operation”
mode and comply with one or more of the following:
1. The robot must STOP when a human is in the collaborative workspace
2. In hand-guiding mode, the robot shall operate at a reduced speed determined
by a risk assessment, but <250mm/sec
3. An appropriate separation distance (IS0 13855 [6]) from the operator shall
be maintained
4. The robot shall be designed to have a maximum dynamic or static powers of
SOW or 150W respectively
5. The robot’s maximum power and force shall be limited to the above
constraints by a control system.

2. Missing standards for new robots

This main problem is that scenarios where robots are designed to have close
human interaction are not included in the standardization process; in particular
cases where a robot has a “human as its workpiece” are not covered. In order to
identify the issues for these new scenarios and environments, it is useful to think
of the different applications using working distances between the robot and
humans as the basis for formulating the requirements. In this respect we can
consider the various cases as “FAR”, “CLOSE’, “TOUCHING’ and
“INVASIVE”. These classifications lead to the following assessment in terms of
their state as regards robot standardization:
FAR: This is the traditional class of industrial robots for which the current
robot standards have been formulated. These standards rely on separating
the robots and humans for safety
CLOSE: The I S 0 10218-1 (2006) [5]industrial robot safety standard covers
these applications as stated above by monitoring and controlling the output
of the robot when the robot is in collaborative mode. It is felt that some new
applications are not adequately addressed; for example, when a mobile
manipulator has to perform a manipulation task while its mobile base is in
motion has not been included in the new safety standard.
TOUCHING: Robots designed to perform tasks that involve touching a
human (continuously or intermittently) are not currently covered in robot
standardization.
700

INVASIVE: Robots designed to perform tasks that involve performing

invasive tasks in a human body are not currently covered in robot
standardization even though a few robotic devices exist and are in full
commercial operation with FDA approval (e.g, da Vinci robot, see
www.intuitivesurgical.com).
To address this shortcoming, IS0 (see www.iso.org) has been investigating
the area to identify what is needed to develop the new standards for the new
robots that are emerging. The Technical Committee that is responsible for robot
standardization is TC 184 under sub-committee SC2. An Advisory Group entitled
Standards for mobile service robots was set up in June 2005 to investigate the
overall standardization requirements and identify the priority area to focus the
initial efforts. The results of the work have been reported in Virk [3] and the
main recommendation was that the safety standard for the new type of robots
should be developed as a matter of urgency. Following this work, and other
recommendations, the need for change within IS0 has been accepted and it has
been decided to broaden the application base for robot standardization and
several decisions have been made:
The “industrial” focus of the IS0 robot committees should be widened to
cover the new non-industrial environments. Details of this widening include:
TC184 is proposing to change it current title, namely “Industrial
automation systems and integration” to “Automation systems and
integration” and its new scope is likely to be: Standardization in the
field of automation systems and the integration of those systems for
design, sourcing, manufacturing and delivery, support, maintenance
and disposal of products and their associated services. This
standardization encompasses the application of multiple technologies,
such as information systems, machines, equipment, robotics for fixed
and mobile robots in industrial and non-industrial environments,
automation & control software, multi-media capabilities, and multi-
modal communications networks. Excluded are base standards in the
following areas, 1) Electrical and electronic equipment as dealt with by
IEC TC 44; 2) PLCs for general application as dealt with by IEC TC
65 and 3 ) multi-media capabilities as dealt with by IEC TClOO.
TC184/SC2 has recently changed its name to “Robots and robotic
devices” from its old name of “Robots for industrial environments”
and its new scope is: Standardization in the field of automatically
controlled, reprogrammable, manipulating robots and robotic devices,
programmable in more than one axis and either fixed in place or
mobile. Exclusions not covered are toys and military applications.
 70 1

A new Project Tsam FT2 (Robots in personal care) has been set up to
develop a new safety standard in the field of robots in personal care
applications, including healthcare applications and excluding entertainment
applications under the chairmanship of Prof GS Virk of CLAWAR Ltd, UK.
A new Advisory Group AG1 (Service robots) has been set up to further
explore needs for standardization in the field of service robots under the
chairmanship of Prof S Moon of Sejong University, Korea.

3. PT2 on Robots in personal care

The new TC184/SC2 Project team PT2 on Robots in personal care (including
healthcare) was set up in June 2006 to develop a safety standard for robots that
are designed to work closely with humans. The committee comprises robot
experts from France, Germany, Hungary, Japan, Korea, Sweden, UK and USA
as well as liaisons with IEEE and IFR, OMG and other IS0 Technical
Committees (TCs). PT2’s work is in its early stages but as it covers a new area
of robot standardization it is important that there are wide ranging discussions,
and views are sought from all relevant sectors on what would be acceptable to
include in the new standard. From the early discussions, within the committee, it
appears that personal care robots can be usefully categorized by some form of
“degree of closeness” as discussed above to determine appropriate safety
measures. It is clear that the “Far” category has been adequately addressed in the
latest IS0 10218 revisions. Hence these standards could also apply to personal
care robots designed to operate far from humans. Therefore, it is only necessary
to focus on the “Close”, “Contact” and “Invasive” categories. Furthermore, I S 0
10218-1 covers some aspects of “Close” operating robots and the cases which
are not adequately handled need to be identified in order that they can be
included in the new standard. This is discussed further next.

3.1. “Close”persona1 care robots not covered in I S 0 10218

Although work is continuing within PT2, some safety issues which are not
covered within the current I S 0 102180-1 standard or nor the planned 10218-2
been identified. Issues that are relevant here include the following:
The operational modes where hands on (master/slave)/handsoff operations,
as well as operations involving unplanned contact need to be addressed.
Proximity sensing measures need to be formulated. These need to cover
human presence, relative positions, velocity and acceleration. In addition,
human body parts sensing will also be needed.
702

Human contacting measures will need to be formulated covering appropriate

forces and pressures that can be imparted on the human.
Measurement procedures for monitoring various human parameters by
robots will also need to be formulated, eg EKG, EMG, EEG
Additional hazards to those considered within I S 0 10218 will also be
needed. These include vibrations and jerking (for repetitive strain), levels
of noise, loss of stability or overturning of robot (falling out or falling on
human), unexpected or unallowed usage (over speeding or overloading).

3.2. Clussifcation of personal care robots for PT2

Personal care (PC) robots cover an extremely wide are and to facilitate the
development of a safety standards it has been decided to classify PC robots into
different categories or robot types. Originally six categories were formulated,
namely, 1) Minimally invasive surgical robots, 2) People carrier robots, 3)
Treatment and examining robots, 4)Rehabilitation robots, 5 ) Physical assistance
robots and 6 ) Mobile manipulators. However, after detailed discussions it was
felt that there was considerable overlap in the groups and hence the categories
have been reduced to the following four:
1. Surgery and medical robots: this group includes all PC robots (non-invasive
and invasive) for monitoring and treatment of persons for achieving and
maintaining good health and providing good healthcare functions.
2. Mobile manipulator robots including the provision of a “zero level service”:
This group covers PC robots that need to move in their environment while
performing specific tasks of manipulation and gripping. The functionality
that is included here is that the robot could actually perform the required
tasks for persons rather than assisting the person to do it. This is referred to
as a “zero level PC service”.
3. Physical assistance robots (including rehabilitation): These robots assist a
person to perform some required tasks. The assistance is aimed at providing
supplementation capabilities only that bring the functionalities of a disabled
injured elderly person to what can be performed by a “normal person”. It is
felt this could be referred to as providing [0-11 level PC service and the “0”
is included as for mobile manipulator robots.
4. People carrier robots (allowing for augmentation capabilities): these robots
are aimed at assisting the mobility of humans. The power requirements can
be large as it may be necessary to provide functionalities that augment
human abilities (eg., moving very fast, lifting heavy loads, etc) are included.
It is felt that this is should focus on providing >1 level PC services but [0-11
level PC services could also be included; in this case clearly there would be
some overlap with physical assistance roots, In view of the large power
 703

outputs that could arise, it is likely that more stricter safety standards may be
needed for these type of robots.

3.3. Issues for PT2 robots

Discussions within the project team have highlighted the issues that needed to be
included in developing the safety standards for FT2 robots. These issues are
important in defining the situations as well as provide the basis on which
appropriate safety requirements could be formulated. The main issues identified
thus far are the following:
1. Definitions & terminology: As the types and uses of robots is growing very
fast it is important that the new situations and scenarios have to be defined
so that the terminology is understood throughout the international
community. In this respect, it is necessary to formulate a variety of
definitions for the terms and terminology (robot types, tasks, usages, etc).
AGl is also likely to recommend that updating the vocabulary is the most
important issue for service robots after safety. FT2 will focus on the terms
and definitions directly to robots in personal care.
2. Classification of PT2 robots on the needs of human: It is clear that the
PT2 robots presented above could be used by different people, namely, (a),
physically impaireddisabled injured people, (b) elderly people and (c)
(normal) able-bodied people to perform supplementary or augmentary tasks.
In addition, the persons could be using the robots in a professional capacity
(this could require some level of competence via formal training of the
professional user as part of his employment) or for personal use (no formal
technical competence can be assumed unless governments decide to
introduce legislation into this area and start asking citizens to train and pass
a “robot operating” test to obtain a robot licence along the lines of a car
driving licence).
3. Classification on closeness level: The “closeness” categories appropriate
for PT2 could be (a) far, (b) close (non-contacting), (c) contacting
(intermittent), (d) contacting (continuous) and (e) invasive. Having such a
fine resolution is useful but the key difference in all these groups is felt to
whether the application is non-invasive or invasive if safety considerations
are the primary concern. Hence all the categories could be usefully grouped
into non-invasive or invasive cases.
4. Classification on the operational environment: The different
environments identified include (a) domestic, (b) industrial/manufacturing,
(c) health and care centres, (d) workplace (office), and (e) public places
(shops, leisure centres, parks, etc.).
It is clear that in formulating the safety standard it will be necessary to
consider all the above issues and adopt a risk analysis approach. The types of
704

hazard and associated protective measures that need to be adopted exist and are
covered in the international standard I S 0 14121-1999 [S]. These include
mechanical hazards (shape, relative location, mass and stability, mass and
velocity, etc), crushing hazard, shearing hazard, cutting or severing hazard,
entanglement hazard, drawing-in or trapping hazard, impact hazard, stabbing
hazard, friction or abrasion hazard, electrical hazards, thermal hazards, noise
hazards, vibration hazards, radiation hazards, hazards due to materials and
substances, hazards due to non-ergonomic designs, combinations of hazards, etc.
Other general guidelines for safety and risk assessment issues relevant to PT2
include IEC GUIDE 51 [9], IEC GUIDE 63 [lo], and IEC GUIDE 73 [ l l ] .
Work to develop the safety standard for personal care robots is progressing
well as reported above and the structure of the new safety standard has also been
discussed. It has been agreed that the new safety standard could be in two parts
along the following lines:
Part 1: Non-invasive personal care robots (including healthcare)
Part 2: Invasive personal care robots (including healthcare).
The initial work will focus on Part 1 and a roadmap for developing it has been
formulated. The details are as follows:
By June 2007 a report on the structure of the Part 1 standard will be presented to
the SC2 plenary meeting. This should contain the main section headings, sub-
sections and possibly some draft contents based on the work carried out by PT2.
A new work item is expected to be proposed at the 2008 plenary meeting of
ISOTC 184/SC2 where a Committee Draft of the Part 1 safety standard should be
available. A Draft standard should be ready for distribution in 2009, followed
by a Final Draft by the 2010 SC2 plenary meeting. The new standard is then
expected to be formally offered for voting and acceptance by the IS0 Members
in 201 1. The main sections of the Part 1 is likely to be along the following lines:
1. Scope
2. Normative References
3. Terms and Definitions
4. Hazard Identification and Risk Assessment
5. Design Requirement and protective measures
6. Information for use
7. Annex A (Normative: list of significant hazards).
In addition to the above work needed for the Part 1 standard, activities will
also be initiated for the Part 2 standard (Invasive personal care robots). For this,
it is vital to ensure good and extensive liaisons are put in place with the relevant
groups involved (especially medical staff and medical equipment
manufacturers). The intention is to organise an event between specialised
medical staff and PT2 members in 2008/09 to highlight the main issues that need
 705

to be included. By the 2009 SC2 plenary committee meeting, it is expected that

a report containing the structure of the Part 2 standard’s sections, sub-sections
and draft content will be prepared. A new work item is then expected to be
made to the 2010 SC2 plenary meeting where a committee draft document
should be available. A similar time-scale as for the Part 1 is then expected to
follow.
Each part will have its own risk assessment based approach which will
utilise IEC Guide 51 [9] and IEC Guide 63 [lo] and it will be essential that
correct links and liaisons are put in place to ensure that each part is developed in
as comprehensive a manner as possible.

4. AG1 on Service robotics

The task of the Advisory Group (AG1) is to explore the needs for
standardisation in the field of service robots so that important areas not already
addressed may be identified and put forward as new work items for ISO. AG1 is
working in parallel with PT2 and meetings have been arranged in co-operation
so that experts can be involved in both committees. The work of AG1 has been
organised into the following four Work Groups:
1. WG1 (Vocabulary): This WG is likely to form the new work item as it is
felt to be the most important after safety. The service robotics area is
requiring new definitions to be made in light of these developments. For
example the current IS0 8373 [7] definition of a robot applies only to
industrial robots and this is not wholly appropriate. This states that an
industrial robot is an automatically controlled, reprogrammable
multipurpose manipulator, programmable in three or more axes, which may
be either jixed in place or mobile for use in industrial automation
applications.
In fact the definition is not adequate when looking at the new scope of
SD2, where robots are seen as “automatically controlled, reprogrammable,
manipulating robots and robotic devices, programmable in more than one
axis and eitherji.xed in place or mobile”. The current trend is to define a
robot as a system that has “motion” and “intelligence” rather than being
“reprogrammable”, “multi-purpose”, etc as defined in I S 0 8373. In view of
this AGl has started to formulate some new definitions that are required.
For example, several definitions of a robot have been produced. These are
as follows:
1. Reprogrammable machine with a degree of autonomy, programmable
in more than one axis, either fixed in place or mobile, and able to
perform task(s). (Note: Degree of autonomy can be from manual
(including tele-operation) to fully autonomous).
706

2. A machine that perceives and manipulates its own state or the state of
its environment under automatic control.
3. Reprogrammable machine with a degree of autonomy which can
purposefully manipulate its own state and the state of its environment
and which is programmable in more than one axis, either fixed in place
or mobile.
4. A reprogrammable machine (or system) with a degree of autonomy (or
intelligence) that perceives and manipulates its own state or the state of
its environment under automatic control.
In addition, two definitions of a service robot have also been formulated;
these are as follows:
1. A robot which provides “usefulness” for the well-being of humans,
society and equipment, excluding manufacturing operations.
2. A robot which provides ‘‘usefulness” for the well-being of humans,
society and equipment, excluding direct manufacturing operations.

2. WG2 (Performance): The WG is aimed at formulating standards to assess

the performance of service robots. It will consider how to objectively assess
the quality of the services that are provided. For example the quality of the
cleaning that a vacuum robot provides is clearly needed. The WG also
includes maintenance issues.
3. WG3 (Software): The WG includes all software issues including
architecture, communication, middleware, HRI and robot-environment
communications.
4. WG4 (Other topics): This WG includes areas such as technical tasks and
modularity.

5. Conclusions
The paper considers the area of standardization for new types of service robots
that are emerging and the current standardization activities that are on-going
under ISO. The paper focuses on PT2’s work on formulating the safety standard
for robots in personal care (including healthcare) and the latest results of the
project team have been presented. This standardization work is vital to remove
the barriers that exist for widening the application base of robots and robotic
systems to new non-industrial environments.

Acknowledgments
The author wishes to acknowledge the valuable contributions to the robot
standardization work reported here by the international experts on the IS0
 707

Project team PT2 on Robots in personal care as well as the Advisory Group AGl
on Service robotics.

References
1 . B. Gates, Dawn of the age of robots, Scientific American (2007).
2. World Robotics 2006, United Nations Economic Commission for Europe,
UNECE and IFR (2006).
3. G.S. Virk, Standards for mobile service robots, Proceedings gthInternational
Conference on Climbing and Walking Robotics (CLAWAR 2006), Brussels,
Belgium, 12-14 September (2006).
4. G.S. Virk, C. Sjostrom, M. Engstrom and W. Trinius, Standards andor
ethics for service robots?, Proceedings Ethics of human interaction with
robotic, bionic, and AI systems: Concepts and policies, Naples, 17-18
October (2006).
5. I S 0 10218-1:2006, Robots for industrial environments - safety
requirements - Part 1: Robot, 34 pages (2006).
6. I S 0 13855:2002, Safety of machinery - Positioning of protective equipment
with respect to the approach speeds of parts of the human body (2002).
7. IS0 8373: 1996, Manipulating industrial robots - Vocabulary (1996).
8. IS0 14121-1999, Safety of machinery - Principles of risk assessment
(1999).
9. IEC GUIDE 51, Safety aspects - Guidelines for their inclusion in
standards (1999).
10 IEC GUIDE 63, Guide to the development and inclusion of safety aspects in
1.

International Standards for medical devices (1999).

11. IEC GUIDE 73, Risk management - Vocabulary - Guidelines for use in
standards (2002).
 PARALLEL PARTICLE SWARM OPTIMIZATION FOR
NETWORKED CLAWAR SYSTEM COOPERATION

FABIO P. BONSIGNORIO
RTD, Heron Robots s.r.l., KR.Ceccardi, 1/18
Genova, I-16121, Italy

This paper shows an example of cooperation between CLAWAR robots based on

particle s w a m optimization. The approach is based on a parallel version of
multiobjective pso and is fit to be implemented on a parallel computing backbone.
The benefits and trade-offs of this approach for managing swarms of CLAWAR systems
are compared and discussed against alternative solutions.

1. Introduction
This paper exposes a model for the cooperation within networks of CLAWAR
autonomous agents, based on particle swarm optimization.
At the same time Reynolds' algorithms are used for flocking.
Particle swarm optimization (PSO) is a population based stochastic optimization
technique first proposed by Eberhart and Kennedy in 1995 [1,2], inspired by the
modeling of bird flocking by Reynolds [ 191.
This method shows some similarities with evolutionary computation techniques
like genetic algorithms . The system is initialized with a population of random
solutions and searches for optima by updating generations. PSO has no, yet
evolution operators like crossover and mutation.
In PSO, the potential solutions, called particles, fly through the problem space by
following the current optimum particles.
Each particle keeps track of its coordinates in the problem space which are
associated with the best solution (fitness) it has achieved so far ('pbest'), the
coordinates in problem space are called 'lbest'. The best value of all the solution
space is called 'gbest'.

The particle swarm optimization algorithm at each time step, changes the
velocity of (accelerating) each particle toward its pbest and lbest locations by
means of a random procedure.
Particle swarm optimization is a simple algorithm that has been proposed for
optimizing a wide range of functions.

708
 709

Nethertheless it was originally proposed to model social behaviour of birds and

other animals.
In this paper we apply a particle swarm optimization method to manage the
cooperation between a group of networked CLAWAR agents with very reduced
computing capabilities, which in turn 'flock' according to a similar algorithm.
From a certain standpoint we apply to the managements of a flock of
autonomous agents the behavior rules supposedly working in natural flocks of
birds following the original example fiom Reynolds and at a different level we
use PSO to optimize the behavior parameters of the single CLAWAR agent.
PSO are comparatively simple algorithms, although they tend to be computing
expensive.
To deal with this fact we have choosen a parallel implementation which can be
used on a cluster of computers or a computing grid.
The usage of a quite similar algorithmic approach for the management of the
agent collective behavior and for the online optimization of the behavior of the
single agent allows a simplification of the computing backbone architecture, '

increasing the robustness of the overall architecture.

2. Flocking algorithm
The flocking algorithm is basicly a parallel version of the one proposed for the
flocking of birds.
Reynolds proposed a simple but effective model to model the behavior of
swarms, such as a flock of birds or school of fish, which are able to move in a
synchronized manner without any central planning or control.
The simulations of bird-like objects flying around, called 'boids', fiom bird-oids,
based on his model looks very realistic and recently has been applied to swarm of
robots [20].
His model, which can considered a multi agent example of 'behavior based'
control strategy, consisted of four behaviors applied to all the 'boids':

1. separation: to make the boids avoid colliding with one another

2. alignment: to make a boid steer in the common direction of his neighbors
3. cohesion: to make a boid steer towards the common position of his
neighbors
4. avoidance: to make a boid steer away fiom other obstacles [Rabin 20021.

In our case the Reynolds' algorithm is applied to a flock of surveillance robot

with a fitness function related to threshold emission in the infrared and sound
simulating an 'intruder', the generalization of the approach is quite easy.

The standard PSO algorithm we use for flocking, implements the following
procedure [ 161:
710

1. Set d equal to the dimension of the fitness function.

2. Create n particles, 0 to n-1, each with a position vector, xi,of dimension d.
3. Assign random position values to each particle
4. For each particle, i :
1. Evaluate fitness (pass position vector values into fitness fimction and
assign return value as fitness)
2. If fitness better than personal best fitness, then assign current position as
personal best position (p3
3. If fitness better than global best fitness, then assign current position as
global best position (pgk)
5 . For each particle, i:

1. Update velocity, v':

( P i- d ) (Pkg - x:)
v;+1 = wv; t Clri t c2r2 (1)
At At

2 . Update position, xi:

6 . If criteria met, end simulation, else repeat from step 4.

The criteria for ending the simulation is usually based on whether the global best
fitness is sufficient, or whether the simulation has run for the maximum amount
of epochs (iterations). It is also important to note that in step 3, the personal best
fitness of each particle, and the global best fitness need to be initialized to very
poor values. At this stage, each particle's velocity is also usually set to zero,
although they may be initialized to random values.

The number of particles, n, inertia weight, w, and the local and global component
variables, 1 and g, are all system parameters of the PSO algorithm.

The local and global component variables, 1 and g, control the influence of the
personal best and global best positions respectively. They are defined as 1 =CIrl
and g = c2r2 , where rl and r2 are random values between 0 and 1, C I and c2 are
acceleration constants that are usually set to values close to .
It has been demonstrated that we convergence if:
 711

The standard algorithm as described above is also called the global best (gbest)
algorithm, because the globally best particle is followed by all the particles. The
local best (lbest) algorithm, however, follows more closely the original model
created by Reynolds. It differs from the gbest algorithm in that the whole swarm
is divided into a number of neighborhoods, where each particle is only aware of
the particles in its neighborhood. There is no global best particle, but rather a
local best particle for each neighborhoods. Each particle is therefore only
influenced by its neighbors, and not the whole swarm. The gbest algorithm can be
seen as the lbest algorithm with one neighborhood consisting of all the particles.
The size of neighborhoods plays an important role in the lbest algorithm. It has
been shown that smaller neighborhoods result in slower convergence, but
generally lead to better results, since a larger part of the search space is explored
and particles are less likely to be trapped in local optima [ 161.

3. PSO behavior optimization

When the flock algorithms is used as an optimization tool there are several
possible improving including dealing with constrained optimization problems,
introducing a craziness operator to increase the likelihood of escaping from a
local minima, and dynamically changing the inertia value w.
The PSO (Particle Swarm Optimization) is a stochastic population based process
depending on the memory of each agent as well as the knowledge gained by the
population as a whole.
The population is called the %warm' while the single agent is called a 'particle'
The craziness operator adds additional randomness to the swarm improving
algorithm's exploitation and reduce the likelihood of premature convergence
and is similar to the mutation operator of a genetic algorithm. Typically a small
subset of particles are selected at each design iteration, and the position andor
velocity vectors are randomly modified.
The inertia parameter w controls the exploration properties of the algorithm. It is
a common practice to start with a larger inertia value (a more global search) that
is dynamically reduced towards the end of the optimization (a more local search).
PSO is a comparatively recent example of non-gradient based, probabilistic
search algorithms. Other examples are evolutionary algorithms and simulated
annealing algorithms.
This class of optimization algorithms have several appealing characteristics.
They are generally easy to implement, can be implemented on a large numbers of
parallel processors, are efficient for finding global or near global solutions, they
712

don't need the computation of derivatives.

The main disadvantage is the weak local search capabilities and the
computational cost.
Thanks to their similarity to evolutionary algorithms, like for example GA
(Genetic Algorihms) they have been applied to design optimization [ 13,14,15]

As told, we base our research on a multi objective particle swarm optimization

algorithm, which is comparatively easy to implement in the form of a parallel
algorithm, in order to make possible to harness the computing capability of a
parallel backbone computing facility.
The multi-objective online optimization of an exapode gait control parameters is
used is perfion. The motivation for using PSO in the present design problem is
the presence of discrete choice variables and numerical noise coming from agent
sensors, for this kind of application a gradient based optimizer would be
comparatively not very effective.
In order to apply the PSO strategy for solving multi-objective optimization
problems, we aim to find a set of different solutions (the so-called Pareto optimal
set).
A new particle replaces its pbest particle usually when this particle is dominated
or if both are incomparable (i.e., they are both nondominated with respect to
each other).
The PSO algorithm is well fit for a coarse-grained parallel implementation on a
parallel or distributed computing network and, in perspective e grid
implementation [21] or on dedicated microprocessors. For each time step (agent
control set iteration), all particles (control set points) are independent of each
other and can be computed in parallel.
The sirect PSO parallel implementation is to simply evaluate the design points
within control set point iteration in parallel, without changing the overall logic of
the algorithm itself. In this implementation, all particles within a design iteration
are sent to the parallel computing environment, and the algorithm waits for all the
analyses to complete before moving to the next design iteration. This
implementation is referred to as a synchronous implementation. This
implementation is not very efficient as it is difficult to obtain that all processors
finish their iteration before the beginning of the next. Moreover we think that an
asyncronous implementation can be more suitable to be didteiubet to an on
demand computing backbone (a 'grid').
The needs of different agents on the environment, by they way, can require
optimization of control set point leading to wery different computation loads.
In order to implement an asynchronous parallel PSO algorithm is necessary to
separate the update actions associated with each point and those associated with
the particle swarm as a whole.
Two operators have to be applied when the iteration is finished for all the swarm
particles: the mutation (crazyness) and the inertia operator.
 713

4. Application scenario

The method have been tested in simulation in Matlab, for algorithm tuning, and
Webots.
Each agent representing a CLAWAR was programmed to evaluate its present
pbest, lbest values, which in our case have a clear interpretation in terms of
physical position of a particle representing the CLAWAR (modelled by a simple
exapode) estimated center of mass position in the 'real' 3d space.
In Webots simulation multibody kinematical additional variable evolution are
simulated.
The considered , semi-structured, scenario is the identification and surrounding
of an intruder in a secured area, the neighborings of a building in an exhibition
area. The robots flock around the intruder, modeled as an infrared and sound
source (see fig. 1)

Building

cr' intruder
\

\
4

vdl d
t
d

Figure 1. Example of agents flocking around an intruder.

\
5. Further Research
We are considering to apply this system in a real world scenario using wheeled
low cost robot appliances and connecting them to a parallel computing backbone
in a first stage and furtherly of newly designed low cost robotic appliances with
legged locomotion.
714

It could also be interesting to blend the PSO procedure appropiately with some
kind of evolutionary algorithms.

6. Conclusions
We have designed a control and coordination system for networked CLAWAR
system for security applications based on a parallel multi objective particle swarm
optimization algorithm and a widely used , in research and even in video game
application, flocking model.
This approach seems very suitable to applications integrating networked robotic
appliances with a parallel computing backbone.

References
1.J. Kennedy, R. C. Eberhart, and Y. Shi,Swarm Intelligence, (2001).
2.5. Kennedy, R. C. Eberhart, “New optimizer using particle swarm
theory,” in Proceedings of the 1995 6th International Symposium on
Micro Machine and Human Science, pp. 39- 43, (1995).
3. J. Kennedy and R. C. Eberhart, “Particle swarm optimization,” in
Proceedings of the 1995 IEEE International Conference on Neural
Networks, vol. 4, (Perth, Australia, IEEE Service Center, Piscataway,
NJ), pp. 1942-1948, (1995).
4. P. J. Angeline, “Evolutionary optimization versus particle swarm
optimization: Philosophy and performance differences,” in Evolutionary
Programming VII, 1998.
5. R. C. Eberhart and Y. Shi, “Comparison between genetic algorithms
and particle swarm optimization,” in Evolutionary Programming VII,
(1998).
7. M. Larvbjerg, T. K. Rasmussen, and T. Krink, “Hybrid particle swarm
optimizer with breeding and subpopulations,” in GECCO 1998: Genetic
and Evolutionary Computation Conference, San Francisco, CA , (1998).
8. B. Al-kazemi, C. K. Mohan, “Multi-phase discrete particle swarm
optimization,” in FEA 2000: Fourth International Workshop on
Frontiers in Evolutionary Algorithms, (2000).
9. R. G. Reynolds and C. J. Chung, “The importance of function
complexity in regulating the amount of information required to guide
self-adaptation in cultural algorithm,” in Proceedings of the Seventh
International Conference on Genetic Algorithms (ICGA97) (T. B”ack,
ed.), San Francisco, CA, Morgan Kauhann, (1997).
10. R. G. Reynolds, “Cultural algorithms: Theory and applications,” in
New Ideas in Optimization (D. Come, M. Dorigo, and F. Glover, eds.),
 715

pp. 367-377, London: McGraw-Hill, (1999).

11.Z. Michalewicz, D. Dasgupta, Evolutionary Algorithms in
Engineering Applications, Springer Verlag, (1997).
12. G.L. Nemhauser, G. L., L.A. Wolsey, Integer and Combinatorial
Optimization, Chapter 3, John Wiley &Sons, (1988).
13. Fourie, P. C. and Groenwold, A. A., The Particle Swarm
Optimization Algorithm in Size and Shape Optimization, Structural and
Multidisciplinary Optimization, 23, pp. 259-267, (2002)
14. Fourie, P. C. and Groenwold, A. A., Particle Swarms in Topology
Optimization, In Proceedings of the Fourth World Congress of
Structural and Multidisciplinary Optimization, Dalian, China, 200 1.
15. G. Venter, G., B. Watson, Efficient Optimization Algorithms for
Parallel Applications, Proceedings of the 8th
AIAA/USAF/NASA/ISSMO Symposium on Multidisciplinary Analysis
and Optimization, AIAA-2000-4819, Long Beach, CA, (2000)
16. A.P. Engelbrecht, Computational Intelligence: An Introduction,
Wiley, (2002).
17. S. Rabin, A1 Game Programming Wisdom, Charles River Media,
2002.
18. F. van den Bergh, An Analysis of Particle Swarm Optimizers, PhD
Thesis, Department of Computer Science, University of Pretoria,
( 2002).
19. C.W.Reynolds, Flocks, Herds, and Schools: A Distributed
Behavioral Mode1,Computer Graphics, 2 1(4), pages 25-34 ,( 1987).
20. J.M. Hereford, “A Distributed Particle Swarm Optimization
Algorithm for Swarm Robotic Applications”, IEEE Congress on
Evolutionary Computation, CEC 2006. Volume, Issue , 16-21, Page(s):
1678 - 1685, (2006).
2 1. F.P. Bonsignorio, “A grid based distributed multiagent enabling
system for intelligent autonomous robot swarms”, ISR 2005,
http://www.isr2005.com/,Tokyo, (2005).
 Performance metrices for improving human-robot interaction

YIANNIS GATSOULIS*t
School of Mechanical Engineering, University of Leeds,
Leeds, West Yorkshire, LSZ SJT, CJK
*E-mail: menigOleeds.ac.uk

GURVINDER SINGH VIRK

School of Engineering and Technology, Massey University
63 Wallace Street, Wellington, New Zealand

The increasing use of robots working together with humans has stressed the
importance of research in human-robot interaction where the robot is seen as a
team member. This paper investigates the human-centric characteristics of sit-
uational awareness, telepresence and workload and how these affect the overall
performance in a task which a human is teleoperating a simulated robot system
in the critical domain of urban search and rescue. We present an overview of
some new methods being developed by the authors together with performance
metrics that can be used to assess the effectiveness of the human-robot inter-
actions proposed. The experimental results, from a set of users consisting of
professional paramedic rescuers, has shown that situation awareness and telep
resence affect to some extent performance. Bivariate regression models were
also shown to better predict performance than simply using a mean and er-
ror model. Workload, which was also hypothesised to affect performance, was
shown to have no effect in the end.

Keywords: Human-robot interaction, teleoperation, performance, situation

awareness, telepresence, workload.

1. Introduction
The increasing use of robots working together with humans in critical appli-
cation domains, such as search and rescue, scientific exploration, etc., has
led to major initiatives to research human-robot interactions (HRI) where
the robots are considered more as members of a team. Moreover, it has been
widely proposed that a human-centric perspective should be adopted in the

t The author has been supported by an EPSRC scholarship.

716
 717

I
Local Environment RemotafOperationaiEnvironment

Fig. 1: Robot teleoperation

design of robotic systems particularly in the research and development of

HRI interfaces.'
Although levels of robot autonomy are increasing in research machines,
tele-operation (Figure 1) is still by far the most common form of robots used
in practice mainly due to reliability and safety reasons. In such situations,
the human operator needs to be provided with high quality information
about the robot and its environment for effective control to be carried out.
In order to utilise a human-centric approach certain human characteristics
that affect task performance (P) should be taken into account, these being
situation awareness (SA), telepresence (TP) and workload (WL). Good SA
can lead to good decision making and hence better performance. However,
given the wide range of definitions and theories,2 more research is needed
to develop a clear consensus on what are the main technical issues and how
these can be solved. TP and WL are two other characteristics that have
been speculated to affect task performance, and despite their long research
their relations to performance are still ~ a g u e .As
~ ?shown
~ in Figure 1 the
main elements are the user, the human-robot interface, the robot and the
local and remote environment, and they are determining SA, TP and WL.
From all these, the system designer has mainly control over the robot and
the human-robot interfaces; to some less extent over the local environment;
and very little over the remote environment and the users.
The very few studies on SA measurement in co-operative intelligent
robots that exist in the literature have focused on specific features of the
robot assisted system, such as the user interaction interface^,^ the use of a
camera,6 autonomous features, e.g. mapping and locali~ation,~ robot health
awareness,* etc. A more comprehensive study is that of J. Riley et aIg>fO
which looks at the overall relations between P, SA, TP and WL, but with
emphasis to TP.
On the same lines of this former study but with emphasis to P, this
718

paper proposes an assessment framework for the evaluation of robot sys-

tems and human-robot interaction interfaces based on measurements and
relations between P with SA, TP and WL. The experimental scenario is a
simulated robotic urban search and rescue mission. The following hypothe-
ses are investigated:
P and SA are positively correlated.
0 P and TP are positively correlated.
P and WL are negatively correlated.
It could be argued that these hypotheses are not new. However, there
have been very few empirical and formal studies including them, and they
suggest further inve~tigation.~Understanding the magnitude of their effect
to performance, as well as developing a prediction model is vital for better
human-centric design and development of robotic systems.

2. Measurement Methods
The response or dependent variable is task performance (P), while the ex-
planatory variables or predictors are situation awareness (SA), telepresence
(TP) and workload (WL).
The experimental task is a computer simulation of a robotic urban
search and rescue scenario. Sixteen subjects were required to teleoperate a
robot and search a floor of a building for any possible casualties, while main-
taining high levels of SA and protecting the robot from any collisions. They
were free to choose any course of action. At the end of the mission or before
the batteries were depleted they were requested to bring the robot into the
exit point. At the end of the trial they were requested to answer to some
questions regarding their SA, TP and WL. All subjects are professional
paramedic rescuers in the Greek health system. Although the majority had
low computer experience, prior training was provided with the system to
ensure that all subjects had a clear understanding of the mission and the
system and were feeling comfortable with using it.
Bivariate correlation analysis (one-sided Pearson’s p ) and bivariate lin-
ear regression model analysis of the forementioned hypotheses were calcu-
lated with the aid of R, a statistical analysis and computing software.ll
The linear model has the form:

y=mx+b (1)
Altough statistical significance is measured at the 95% level, the 90%
level is also reported, and hence a statistical insignificant result is reported
 719

on this latter one.

2.1. Performance measurement

Performance is measured according to Equation 2a:
AC+RS+EP
PK
t
where,
P : performance score, P E R(1.00,. . . ,100.00)
AC: percentage of the area covered
RS: a reward for bringing the robot safely back
E P : exceptional performance, a reward for extensive use of robot, awarded
if the percentage of the area covered is more than a threshold value
t: mission time
In contrast to the RobocupRescue competition where task performance
measurement is mainly based upon the number of victims and their diffi-
cult in being located,12 we preferred the area covered over time as the main
variable affecting performance as a more objective measure. The reason is
that based on the assumption that the more area searched, the more likely
it is to locate any existing casualties, or in the case than none exists time
can be saved as the area can be marked as “clear”. Also number of casu-
alties located and the level of difficulty in finding them is an inappropriate
measurement in this case as the casualties are placed in clearly visible open
spaces without being hidden behind any obstacles. As a result the casualties
are relatively easy to identify once seen in the camera. Furthermore, they
do not provide any additional visual (e.g. thermal signature) or audio cues.
This actually means that locating a victim is merely a function of visiting
the area close to the casualty. For this reason, the area covered is a more
objective measurement of performance.

2.2. Situation Awareness Measurement

Situation awareness has mainly been investigated in the area of avionics
and air traffic management. However, the majority of these methods are
either too specific or too simple to be used in the domain of robotics. For
this reason we developed a post-trial self-rating questionnaire to address
the needs of SA in a robot tele-operation scenario. The questions have been
~

aThe actual equation is a bit more complex involving some normalising constants, and
is beyond the scope of the paper.
720

designed to cover the different dimensions of the user’s SA, such as local-
isation awareness, time awareness, coverage awareness, spatial awareness,
etc.l3,l4 along the three levels of SA as proposed by M. Endsley,15 namely
having a clear perception of the available data, comprehension of the data
to meaningful information and being able to make good predictions of fu-
ture states. Example questions of the measurement methods are shown in
Table 1. The user is asked to rate herself on a 6-point scale.

Table 1:Example questions of the situation awareness measurement method

-How easy was it to avoid the hazards (e.g. narrow doors)?

-How easy was it to keep track of the status of the various modules of the robot
-How easy was it to keep track of the time?
-How easy was it to predict what would happen next?

2.3. Telepresence Measurement

Telepresence is measured after the trial through a self-rating questionnaire
developed by Witmer and Singer,lGwhich was modified to meet the re-
quirements of our search and rescue experimental task. Example questions
are shown in Table 2. The user is asked to rate herself on each question on
a 7-point scale.

Table 2: Example questions of the Witmer-Singer presence questionnaire

-How much were you able to control the various events?

-How responsive was the environment to actions that you initiated?
-How aware were you of the events occurring in the real world around you?
-How often do you player video/computer games?

2.4. Workload Measurement

Workload is measured after the trial through NASA-TLX,I7 which is a
multi-dimensional weighted rating method. The different dimensions are
mental demand, physical demand, time demand, performance, effort, stress
and frustration. From those the physical demand dimension was omitted
as it is inappropriate for this computer simulation task.
 72 1

3. Results
The Lilliefors normality test scores were for P, D(16) = 0 . 1 5 , ~> .05;
for SA, D(16) = 0 . 1 2 , ~> .05; for TP, D(16) = 0 . 1 2 , ~> .05; and for
workload WL, D(16) = 0 . 0 9 , ~> .05; indicating that all samples are similar
to a normal distribution, and hence allowing parametric statistic tests to
be used. Descriptive statistics of the distribution for each dataset of each
variable are shown in Table 3.

Table 3: Descriptive statistics of the distribution of the dataset of each

variable
~ ~ ~ ~ ~ ~~

variable mean u min max valid-range

Performance 51.25 17.82 19.44 89.01 [l- 1001

Situation Awareness 4.24 0.83 2.82 6.00 [1 - 61
Telepresence 4.44 0.37 3.78 5.13 [1- 71
Workload 65.95 13.53 44.00 87.00 [l - 1001

Scatterplots of P/SA, P/TP, and P/WL are shown in Figure 2. The

term “outlier” is used in the context in which a case seems to largely deviate
from the cluster of the rest of the values. Correlation and linear regression
model analysis were carried out for both the complete set of values ( N = 16)
as well as when omitting these outliers. In each case the sample size is
reported to signify whether the results are for the complete set or the filtered
one.
The results of the correlation analysis for the complete set show that
there is a significant positive correlation between P and SA ( N = 16, p =
.57, p < .05). This positive correlation is even stronger when using the
filtered dataset ( N = 15, p = .71, p < .01). A significant positive corre-
lation also appears between P and TP ( N = 16, p = .47, p < .05). For
the complete set it seems that there is no relationship between P and WL
( N = 16, p = .09, p > .1), while for the filtered dataset a small positive
correlation appears to exist (N = 14, p = .34, p > .l);this is also verified
by the wide spread of points in the scatterplot shown in Figure 2.(c). Both
results are not statistically significant, not even at the 90% level. Table 4
summarises all these results.
In the case of a bivariate linear regression model of PISA, SA can
account for 31.0% of the variation of P in the complete dataset ( N =
16, R2 = .31), while in the filtered dataset this increases to 50% (N =
722

8 - 0
. Normal
Ovnler
8
8
-
-
0 Normal

co- L-
8 8 -
g 8-
0

0
0 8
0 0
e
i 8
8;
0 0
0 0 0

3 - o
s- 0 0

8 -
P- I
. 4 -
8 ,
0

8 .
R
50 M) 70 m
Wohlmd

(c) Workload

Fig. 2: Scatterplots of Performance with

Table 4:Bivariate correlation, Pearson's p, one-sided (P: Performance, SA:

Situation Awareness, TP: Telepresence, WL: Workload)

P/TP 16 .47** .22

16 .09 .01
14 .34 .ll
**significant at the .05 level
***significant at the .01level

15, R2 = .50).In both cases the linear model predicts P significantly better
than the mean value model ( N = 16, F = 6.30,p < .05;N = 15, F =
13.14,p < .05).A bivariate linear regression model analysis of P / T P shows
that TP can account for 22% of the variation of P ( N = 16, R2 = .22)
 723

and is (significant at 90% level) better than the mean model (N = 16, F =
4.04, p < .1). For P / W L , a bivariate linear regression model analysis shows
that WL has no or very little effect on the variation of P and the linear
model is no better to the mean model for the complete dataset (N =
16, R2 = .01, F = 0.12, p > . l ) or the filtered one (N = 14, R2 =
.11, F = 1.55, p > .1). These results are summarized in Table 5 , while
Table 6 shows the bivariate linear model coefficients.

Table 5: Summary of bivariate linear regressions

Y/X N R R2 adj R2 F
16 .57 .31 .26 6.30"
15 .71 .50 .46 13.14"

PITP 16 .47 .22 .17 4.04*

16 .09 .01 -.06 .12
14 .34 .11 .04 1.55
*significant at the .1 level
**significant at the .05 level
***significant at the .01 level

Table 6: Coefficients of bivariate linear regressions (y = ma: + b)

Y/X N b SEb tb m SE, tm
16 0.25 20.67 .01 12.03 4.79 2.51"
15 -3.82 16.07 -.24 13.58 3.75 3.62""
P/TP 16 -51.05 51.05 -1.00 23.05 11.47 2.01'

16 43.27 23.57 1.84' .12 .35 .35

14 11.49 31.02 .37 .55 .44 1.24
*significant at the .1 level
**significant a t the .05 level
***significant at the .01 level

4. Discussion
The experimental results have verified the hypotheses that the human-
centric characteristics of situation awareness (SA) and telepresence (TP) are
724

two main factors that affect performance (P) in a positive related manner,
and should be taken into account when designing robot systems and human-
robot interaction interfaces. For the complete datasets they both together
account for 53% of the variance of P, while the filtered datasets show an even
higher contribution (SO%)b. The hypothesis that P and workload (WL) are
negatively correlated should be rejected as it appears to be no correlation
at all in the studies carried out. However, this can be explained by the fact
that only very low and very high levels of WL have a negative impact on
P. In the former case the user may have not allocated sufficient attentional
resources while in the latter it might be beyond the person’s processing
capabilities. On the other hand, in moderate levels of WL the user is kept
occupied without being overloaded. The task difficulty of this particular
experimantal scenario can be considered as easy to intermediate, with the
values of WL verifying thisc (mean = 63.95, ~7= 13.53), and this can be
the reason why workload seems to have no relation to P.
Bivariate analysis has the limitation that it investigates performance
and a variable without limiting the effect of the rest. A more accurate anal-
ysis and improvement is semi-partial correlation and multiple regression,
as well as partial correlation to see if there are any relations between the
explanatory variables themselves. It would also be interesting to further
explore which particular dimensions of SA, which are already measured by
the assessment technique, have a bigger impact to the variance of P, as SA
in this analysis was treated as an end product.

5. Conclusions
This study has shown that the human-centric characteristics of situation
awareness and telepresence affect task performance to some extent. These
are within the sphere of influence of a system designer unlike remaining vari-
ables that might influence performance such as for example task difficulty,
training, experience and cognitive capabilities of the user. By taking into
account these and by using an assessment framework as the one proposed,
system designers can compare their developed systems and human-robot
interaction interfaces, as well as being able to predict to some extent their
task performance. By further analysing the individual dimensions of each

bR2 = .30, F = 5.53, p < .05 of P l T P calculated for the same dataset as the filtered
one used for P I S A
=Although it has to be mentioned that the subjects had not experienced very easy or
very difficult scenarios, something that may have influenced these results.
 725

variable, system designers can focus on particular aspects of the overall sys-
tem t h a t support th e ones t hat have a bigger impact, resulting in a more
efficient research and development.

References
1. J. L. Burke, R. R. Murphy, E. Rogers, V. J. Lumelsky and J. Scholtz, IEEE
Transactions on Systems, Man and Cybernetics 34 (2004).
2. C. Dominguez, Can SA be defined?, in Situation Awareness: Papers and
Annotated Bibliography Interim Report No. AL/CF-TR-I994-0085, eds.
M. Vidulich, C. Dominguez, E. Vogel and G. Mcmillan 1994 pp. 5-15.
3. T. B. Sheridan, Presence 1,120 (1992).
4. P. Tsang and G. F. Wilson, Mental workload, in Handbook of Human Factors
and Ergonomics, ed. G. Salvendy 1997 pp. 417-449, 2nd edn.
5. J. C. Scholtz, B. Antonishek and Young, IEEE Transactions on Systems,
Man and Cybernetics-Part A : Systems and Humans 35 (2005).
6. S. B. Hughes and M. Lewis, IEEE Transactions on Systems, Man and Cy-
bernetics - Part A: Systems and Humans 35,513 (2005).
7. H. A. Yanco and J. Drury, “Where am I?” Acquiring situation awareness
using a remote robot platform, in Proc. of the IEEE Conference on Systems,
Man and Cybernetics, Oct. 2004.
8. K. M. Reichard, Robotics and Autonomous Systems, 105 (2004).
9. J. M. Riley, The utility of measures of attention and situation awareness for
quantifying telepresence, PhD thesis, Department of Industrial Engineering,
Missisipi State University2001.
10. J. M. Riley, D. B. Kaber and J. V. Draper, Human Factors and Ergonomics
in Manufacturing 14,51 (2004).
11. R Development Core Team, R: A Language and Environment f o r Statistical
Computing. R Foundation for Statistical Computing, Vienna, Austria, (2007).
12. A. Jacoff, B. Weiss, E. Messina, S. Tadokoro and Y. Nakagawa, Test arenas
and performance metrics for USAR robots, in Proc. of the 2003 IEEE/RSJ
International Conference on Intelligent Robots and Systems, (Las Vegas,
2003).
13. Y. Gatsoulis and G. S. Virk, Modular situational awareness for CLAWAR
robots, in Proc. of CLAWAR 2005, 8th International Conference on Climbing
and Walking Robots, 2005.
14. Y. Gatsoulis, G. S. Virk, M. Parack and A. Kherada, “What’s going on?”
an alternative approach into investigating human-robot interactions, in Proc.
of CLAWAR 2006, 9th International Conference on Climbing and Walking
Robots, 2006.
15. M. R. Endsley, Human Factors 37,32 (1995).
16. B. G. Witmer and M. J. Singer, Presence: Teleoperators and Virtual Envi-
ronments 7,225(June 1998).
17. S. G. Hart and L. E. Stavenland, Development of NASA-TLX: Results of
empirical and theoretical research, in Human Mental Workload, eds. P. A.
Hancock and N. Meshkati 1988 pp. 139-183.
 REAL-TIME COMPUTATIONAL COMPLEXITY OF THE
ALGORITHMS FOR A SINGLE LINK MANIPULATOR SYSTEM

M.A. HOSSAIN'
'Department of Computing, University of Bradford

M.N.H. SIDDIQUE'
'School of Computing and Intelligent Systems, University of Ulster

M.O. TOKH13 AND M.S. ALAM4

'*4Departmentof Automatic Control and Systems Engineering, University of Shefield

This paper presents an investigation into the real-time computational complexity of the
algorithms for a single link manipulator system. A dynamic simulation algorithm of a
single link manipulator system using finite difference (FD) method is considered to
demonstrate the real-time computational complexity. The simulation algorithm of the
manipulator system is analyzed, designed in various forms and implemented to explore
the impact in implementing real-time. Finally, a comparative real-time computing
performance of various forms of the new and previously reported algorithms is presented
and discussed to demonstrate the merits of different design mechanisms through a set of
experiments.

1. Introduction
Although computer architectures incorporates fast processing hardware
resources, high performance real-time implementation of a complex algorithm
requires an efficient design and software coding of the algorithm so as to exploit
special features of the hardware and avoid associated shortcomings of the
architecture. In practice, more than one algorithm exists for solving a specific
problem. Depending on the formulation, each can be evaluated numerically in
different ways. As computer arithmetic is of finite accuracy, different results can
evolve, depending on the algorithm used and the way it is evaluated. On the
other hand, the same computing domain could offer different performances due
to variation in the algorithm design and in turn, source code implementation. The
choice of the best algorithm for a given problem and for a specific computer is a
difficult task and depends on many factors, for instance, data and control

726
 727

dependencies of the algorithm, regularity and granularity of the algorithm and

architectural features of the computer domain [l], [ 2 ] .
The ideal performance of a computer system demands a perfect match between
machine capability and program behaviour. Program performance is the
turnaround time, which includes disk and memory accesses, input and output
activities, compilation time, operating system overhead, and central processing
unit (CPU) time. In order to shorten the turnaround time, one can reduce all
these time factors. Minimising the run-time memory management, efficient
partitioning and mapping of the program, and selecting an efficient compiler for
specific computational demands, could enhance the performance. Compilers
have a significant impact on the performance of the system [3]. This means that
some high-level languages have advantages in certain computational domains,
and some have advantages in other domains. The compiler itself is critical to the
performance of the system as the mechanism and efficiency of taking a high-
level description of the application and transforming it into a hardware
dependent implementation differs from compiler to compiler [31.
The performance demand in modern real-time signal processing and control
applications has motivated the development of advanced special-purpose and
general-purpose hardware architectures. However, the developments within the
software domain have not been at the same pace and/or level as within the
hardware domain. Thus, although advanced computing hardware with significant
levels of capability is available in the market, these capabilities are not fully
utilised and exploited at the software level. Efficient software coding is essential
in order to exploit the special hardware features and avoid associated
shortcomings of the architecture. There has been a substantial amount of effort
devoted to this area of research over the last decade 141, 151, [61, [71, [SI.
This paper presents an investigation into the algorithm analysis, design and
software coding for real-time computational complexity. It is worth mentioning
that this is an extension of the research reported earlier in [9]. Three new
analyses and design models are explored, implemented, tested and verified to
demonstrate the computational complexity of the algorithms of a manipulator
system implemented in real-time. These new design approaches are then
compared with the previous three approaches to demonstrate the merits and
capabilities of the algorithms.

2. The Single Link Manipulator System

Figure 1 shows a schematic representation of a single-link manipulator system. A
control torque z is applied at the pinned end (hub) of the arm by an actuator
728

motor. 8 represents the hub angle, POQ is the original co-ordinate system
(stationary coordinate) while P‘OQ’ is the co-ordinate system after an angular
rotation 8 (moving coordinate). 1, is the inertia at the hub, I , is the inertia
associated with a payload M, at the end-point and u is the flexible
displacement (deflection) of a point at a distance x from the hub. The dynamic
equation of the flexible manipulator, considered as an Euler-Bernoulli beam
equation, can be expressed as [ 101:

where, y ( x , t ) is the displacement (deflection) of the manipulator at a distance

n from the hub at time t , p is the density per unit length of the manipulator
material, E is Young modulus, I is the second moment of inertia, z(x,t) is the
applied torque. The product El represents the flexural rigidity of the
manipulator.

Figure 1. Schematic representation of the flexible manipulator system

Discretising the manipulator in time and length using central finite difference
methods, a discrete approximation to equation (1) can be obtained that can be
stated in matrix notation as [lo]:
 729

where Yj,j+l is the displacement of grid points i = 1,2; -.,n of the manipulator
at time step j + l , Y i , j and Yi,j-l are the corresponding displacements at time
steps j and j-I respectively. A and B are constant n x n matrices whose entries
depend on the flexible manipulator specification and the number of sections the
manipulator is divided into, C is a constant matrix related to the given input
torque and F is an n x 1 matrix related to the time step At and mass per unit
length of the flexible manipulator.

3. Algorithm Design
The FD simulation algorithm of the manipulator system as shown in equation 2
is considered to demonstrate the different design impact in implementing the
algorithm in real-time. As mentioned earlier, three different approaches have
been reported earlier [9]. This section presents three new design approaches of
the algorithm with their advantages and disadvantages.

Algorithm4 The ‘Algorithm-4’ is listed in Figure 2, in which the new

methods of Algorithm4 were applied with the concepts of Algorithm-3
reported in [9]. Three distinct calculation runs are performed during each
iteration, but instead of listing the instructions for each segment separately,
nested loops are used to limit the number of instructions (source code lines) in
the main program loop. The benefits of employing this technique are identical
with those listed in the description of Algorithm-3. However, it possesses the
same disadvantage of overhead produced by the complex substitutions required.

Initialise the parameters of the Manipulator.

Step 1: Calculate Current position of different segments y[i][2]; where i=O-19;
LOOP {
yor21= (A[ 131lI*YO[ 11+ A[ 11P I *Y 1[ 11 + A[ 11[31*Y2[1I>- (B[ 11[ ~l*Yo[ol+
B [ 11[2]*yl [O]) + tauu]*C;
y[11[21= (A[21[1I*Y[Ol[~l+A[21[2l*Y[ll[ 1l+A[21[31*y[21r11+~[21[41*y[31[11)-
~ ~ ~ ~ l ~ ~ l * Y ~ ll[Ol ~ l +~B[21[31*y[21[01);
~ l + ~ ~ ~ l ~ ~ l * Y ~
LOOP (
y[i] [2]=(A[i+l][i-l]*y[i-2][ l]+A[i+ 11[3]*y[i][ l]+A[i+l] [i+2]*y[i+ 11[i]
+A[i+ l][i+3]*y[i+2][i])-(B[i+l][i]*y[i-l][O]+B[i+l] [i+l]*y[i] [O]
+B [i+ 1I [i+21*y[i+ 1I [O]);
1
730

//Step 2 : Calculate y[i][O] ; where i = 0-19 ;

YO[OI=(A[ll[l1*~0[21+A[11[21*~1[21+A[11[31*~2[21)-(B[11[11*~0[~1+
B[ 1][2]*yl[ 11) + tauu]*C;
Y[ll[OI= ("I[ 11*~[01[11+ 4 2 1[ ~ I * Y11[[ 11+A[21[31*~[21~
1l+A[21[4I*y[31[I])-
("][ ll*Y [OI [OI+B[21[21*Y[ 11[OI + B[21[31*Y[21[Ol);
LOOPI
y[i] [O]=(A[i+l][i- 1]*y[i-2][ l]+A[i+ I] [3]*y[i][ l]+A[i+l][i+2]*y[i+ 11[i]
+A[i+ 11[i+3]*y[i+2] [i])-(B [i+l][i]*y[i- 11[O]+B[i+l][i+l]*y[i] [O]
+B[i+l][i+2]*y[i+ 11[O]);
1
y[ 18][0]=(A[19][ 17]*y[16][1]+A[19][18]*y[ 17][1]+A[19][19]*y[ 18][ 1]+A[ 19][
201 *Y [ 191[ 11)- (B[ 191[181*Y[ 171[OI + B[ 191[W*Y[ 181[01+
BC 191[201*Y[ 191PI);
y19[0]= (A[20][18]*y17[2]+ A[20][19]*y18[2] + A[20][20]*y19[2])-
B [20][20]*y 19[11;

// Step 3 : Calculate y[i][l] ; where i = 0 -19.

YO[lI=(A[ 11[~1*~0[01+ [OI + A[ 11[31*~2[01)-(B[ 11[11*~0[21+
A[ 11[21*~1
B [ 13[2]*y 1[2]) + tauti] *C;
Y [I1[ 11= ("11 11*y[Ol[11+ A[21[21*~[11[11+A[21[31*~[21[ 1I+A[21[41*~[31[11)-
(BPI [ 11*y[oI[Ol+BP I [ ~ I * Y1l[Ol
[ + B[21[31*~[21[01);
LOOP {
y[i] [ l]=(A[i+ l][i- l]*y[i-2][l]+A[i+ 1][3]*y[i][ l]+A[i+ l][i+2]*y[i+l][i]
+A[i+ 13[i+3]*y[i+2] [i])-(B[i+l][il *y[i-ll[O]+B[i+ l][i+ll*y[il [OI
+B[i+ l][i+2]*y[i+l][O]);
1
y[ 18][l]=(A[ 19][ 17]*y[16][1]+A[19][18]*y[17][ 1]+A[ 19][19]*y[ IS][ 1]+A[19][
201 *~[191[11)- (B [ 191[ 181*Y[171[OI + B[ 191[191*y[181[01+
B[ 191[20l*Y[19"l);
y19[1]= (A[20][18]*y17[0]+ A[20][19]*y18[0] + A[20][20]*y19[0])-
B[201[201*y19[21; 1
..........................................................................................

Algorithm-5: The 'Algorithm-5' is listed in Figure 3. This makes use of the

fact that access to the oldest time segment is only necessary during re-calculation
of the same segment. Hence, it can directly be overwritten with the new value as
 73 1

shown in Fig. 4. The conventional re-calculation algorithm in Fig. 2 requires

three memory segments in the time domain. In contrast, Algorithm-5 is
optimized for the particular discrete mathematical approximation of the
governing physical formula, exploiting the previously observed features.
It is noted that this particular algorithm is not suitable for applications for
which the previous assumption does not hold. This technique may offer a major
performance advantage over the conventional rotation method, in particular
when the number of segments is increased.
732

I I

II I
Calculations on registers
I
I

Figure 4.Recalculation with reduced memory allocation

Algorithm-6: Algorithm-6, as shown in Fig. 5 , is based on improvements

achieved with Algorithm-6. Additionally, the notion of nested loops was
incorporated. The advantages and disadvantages of this approach were identified
earlier and remain true for this particular algorithm.

Figure 5. Algorithm-6
Initialise the parameters of the Manipulator.
LOOP (
Step 1: Calculate Current position of different segments y[i][ 13; where i=O-19;
YO[ 11= (A[ 1I [ 11*Yo[ 11+ A[ 1"l*Y 1[ 11 + A[ 11131*Y2[ 11)- (B[ 11111*Y0[01+
B[1][2l*y1[0]) + taulj]*C;
Y[ 1I[ 1I= "1( [ 11*Y[OI [ 11+ 1" 121*Y[ 11[I]+"] [31*Y[21[ 11+"1[41 *Y[31[11)-
(BPI[ 1 l * ~ ~ ~ l [ O l + BP [I2*Y
l [ 1l[Ol + B[21PI*y [21[01>;
LOOP I
y[i] [ l]=(A[i+ 11[i- 11*y[i-2][ l]+A[i+1][3] *y[i][ l]+A[i+l] [i+2]*y[i+ 1][ 11
+A[i+ 1I [i+3l*y[i+2][ l])-(B[i+ll[i] *y[i- 11[Ol+B[i+11[i+ 11*y[il[Ol
+B[i+ lI[i+2]*y[i+ 13[O]);
1
y[ IS][ l]=(A[ 19][17]*y[16][ 1]+A[ 19][18]*y[ 17][1]+A[19][19]*~[18][1]+A[19][
201 *y[ 191[13)- (B[ 191[181*Y[ 17"l + B[ 191[19I*Y[l81P I +
B t 191P O I *Y [ 191[OD;
y19[1]= (A[20][18]*y17[1]+ A[20][19]*y18[1] + A[20][20]*y19[1])-
B [20][20]*y19[01;
 733

LOOP {
y[i][O]=(A[i+l][i-l]*y[i-2][l]+A[i+l][3]*y[i][ l]+A[i+l][i+2]*y[i+l J[ 11
+A[i+ ll[i+3l*y[i+2][ 11)-(B[i+ I][i] *y[i-11[O]+B[i+11[i+ 11*y[i][O]
+B[i+l][i+2]*y[i+ 1][0]);
1
y[18][0]=(A[ 19][ 17]*y[16][1]+A[19][ 18]*y[17][1]+A[19][19]*y[ IS][ l]+A[ 19][
201 *Y[191[ 11)- (B[ 191[ 181*Y[171[01+ B[ 191[19l*y[181[01+
B[ 191[201*Y[191[01);
y[ 191[OI= (A[201[ 18l*y[171[1I+ A[201[ 191*~[181[11+AD01P O I *y[ 191[11)-
BC201[201*y[191[01; 1

4. Experiments and Results

Three new design methods were implemented for similar specification to
demonstrate the merits and capabilities of the approaches. Same compiler (C++)
and computing domain is used to implement the simulation environment for
testing and verification. It is worth noting that a fixed number of segments for
various iterations were considered in implementing all the algorithms for the
sake of consistence. Moreover, the sampling time 0.000217904 sec is considered
for each iteration, therefore, required real-time performance of 5000 iterations,
as an example, should be 5OOOXO.OOO217904= 1.0895sec.
Figure 6 shows the comparative performance of the three new algorithms for
20 segments. It is observed that the execution time for the algorithms increases
almost linearly with the increment of iterations. It is also noted that the
Algorithm-5 performs best among the algorithms. In contrast, Algorithm-6
performs waste among the three algorithms. It is also observed that Algorithm4
and Algorithm-6 have not achieved required performance to implement in real-
time.
Table 1 demonstrates the relative performance of the six algorithms, including
previously reported [9] three. It is worth mentioning that the performance in
Table 1 is presented as ratio of the Algorithm-1, Algorithm-3, Algorithm-4,
Algorithm-5 and Algorithm-6 relative to the Algorithm-2 to keep consistence of
the previous report [9]. It is observed that the Algorithm-4, Algorithm-5 and
Algorithm-6 offer relatively better performance as compared to the Algorith-1
and Algorim-3. On the other hand, the Algorithm-5 is about 1.21 times (average)
slower as compared to the Algorithm-2. Execution time of the Algorithm-6 is
more then double as compared to the Algorithm5 and very similar level of
Algorithm-4. Thus, the design mechanism employed in Algorithm-2 and
Algorithm-5 can offer potential advantages for real-time implementation.
734

1 2 3
Number of itlsralians ( XI0000 )

Figure 6. Performance comparison of the various forms of the Algorithms and

real-time requirement (WQ)

Table 1: Performance of the Algorithms relative to the Algorithm-2 (Alg-2)

5. Concluding Remarks
This paper has presented an extension of the investigation reported in [S].The
investigation has focused into the algorithms analysis, design, software coding
and implementation so as to reduce the execution time and, in turn, enhance the
real-time performance. Three new design approaches for real-time
implementation have been proposed and demonstrated experimentally. It has
been observed that the execution time and in turn, performance of the algorithms
varies with different approaches in a real-time implementation context. It is also
noted that only Algorithm-2 (previously reported) and Algorithm-5 have
 735

achieved real-time performance for various number of iterations. Although, the

other four algorithms can also be implemented in real-time by using high
performance computing domain, however, identification of the suitability of
Algorithm design and implementation mechanism for best performance is a
challenge, in particular for a specific architecture.

References
1. A. U. Thoeni, Programming real-time multicomputers for signal processing,
Prentice-Hall, Hertfordshire, UK (1994).
2. M. 0. Tokhi and M. A. Hossain, CISC, RISC and DSP processors in real-time
signal processing and control, Journal of Microprocessors and Microsystems,
19(5), UK. pp. 291-300 (1995).
3. G. Bader and E. Gehrke, On the performance of transputer networks for solving
linear systems of equation, Parallel Computing, 1991, 17, pp. 1397-1407 (1991).
4. M. 0. Tokhi, M. A. Hossain, M. J. Baxter and P. J. Fleming, Heterogeneous and
homogeneous parallel architectures for real-time active vibration control, IEE
Proceedings-D: Control Theory and Applications, 142, (6), pp. 1-8 (1995).
5. B. N. Bershad and D. Lee, T. H. Romer and B. J. Chen, Avoiding conflict misses
dynamically in large direct-mapped caches, Proceedings of Sixth International
Conference on Architectural Support for Programming Languages and Operating
Systems (ASPLOSP4), San Jose, Canada, pp. 158-170 (1994).
6. B. Clader, C. Krintz, S . John and T. Austin, Cache -concious data placement,
Proceedings o f Eighth International Conference on Architectural Support for
Programming Languages and Operating Systems (ASPLOS'98), San Jose, Canada,
pp. 139-149 (1998).
7. K. Hwang, Advanced computer architecture - Parallelism, scalability,
programmability, McGraw-Hill, California, USA (1993).
8. M. A. Hossain, U. Kabir and M. 0. Tokhi, Impact of data dependencies for real-
time high perjormance computing, Journal of Microprocessors and Microsystems,
26(6), pp. 253 - 261 (2002).
9. M. A. Hossain, M. N. Siddique, M. 0. Tokhi and M. S. Alam, Design
Constraints in Implementing Real-time Algorithms for a Flexible Manipulator
System, CLAWAR 2005: CIimbing and Walking Robots, M 0 Tokhi, G S Virk
and M A Hossain (eds.), Springer, Germany, pp. 583-590 (2005).
10. A. K. M. Azad, Analysis and design of control mechanisms for flexible
manipulator systems, PhD Thesis. Department of Automatic Control and
Systems Engineering, The University of Sheffield, UK (1994).
 SOFTWARE AND COMMUNICATIONINFRASTRUCTURE
DESIGN OF THE HUMANOID ROBOT RH-1

D. KAYNOV, M. ARBULU, P. STAROVEROV, L. CABAS, C. BALAGUER

Robotics Lab,
Department of Systems Engineering and Automation,
University Carlos III of Madrid, Av. Universidad 30,
Legan& Madrid, Spain

This paper presents an advanced control system for the humanoid robot. The main
advantage of the proposed control architecture is wide use of the standardized and
frequently used in the automation industry solutions. It provides scalability, modularity
and the application of standardized interfaces and brings the design of the complex
control system of the humanoid robot from a closed laboratory to the industry. The main
parts of the proposed software control system architecture and communication
infrastructure are presented. As well as some aspects of the humanoid robot walking
control from the automation side are discussed. The designed software control system
was implemented with the Rh-1 humanoid platform, the second phase of the Rh project,
which have been launched by the Robotics Lab at the University Carlos 111 of Madrid at
2002.

1. Introduction
Recently the development of sophisticated humanoid robots has increased and it
has become very active research area. There is growing interest not only in
academic area but in some industrial areas too. Several humanoid robots have
been developed in these years. One of the best humanoid robots is HRP-2
designed by Kawada Industries [I]. The most impressive humanoid robot
should be ASIMO constructed by HONDA [2]. It is necessary to mention that
the great success of ASIMO makes the current research of humanoid robots to
become very promising working field for scientists and engineers. One of the
recently presented successful projects is KHR-2 humanoid robot constructed by
KAIST (Korea Advanced Institute of Science and Technology) [3].
In the major part of works on the humanoid robot problem, the design is
impartially concentrated on the walking problem. The software architectures
description is superficial and does not give the clear idea about criterions used
to design the system and its validity. The goal of this paper is to determine the

736
 737

humanoid robot software systems as the individual type of the architecture

inside the general control problem.
This paper presents the Rh-1 robot, introduces its communication
~ f r a s t ~ c t uand
r e software control architecture. Some aspects of the humanoid
robot control from the automation side are also discussed.

2, H u ~ a n o i drobot Rh-1
Rh-1 is a new humanoid robot in the second phase of the Rh project. The robot
is developed for several principal tasks such as human care, maintenance of
dangerous for human health plants, entertainment etc Figure 1 shows

Figure 1 Mechanical design and developed Rh-1 robot,

Electrical design of Rh-1 robot is based on distributed motion control

philosophy there each control node is an independent agent on the network.
Figure 2 shows the physical hstribution of the hardware inside the humanoid
robot.

Figure 2 Hardware distribution inside the humanoid robot.

738

The mechanical and dynamical design of Rh-l is completed. The detail

software and communication infrastructure design will be presented in
followings sections.

3. Software Architecture
A humanoid robot can be considered as a plant were the shop floor consists of
a series of cells (intelligent motion controllers and a sensors) managed by
controllers (Main Controller, Communication Supervisory Controller, etc.) In
general, there are two basic control tasks for the control system of a humanoid
robot. The first goal is to control all automation and supervise the data
transmission. And the second goal is to control and monitor the entire floor to
detect failures as early as possible and to report on performance indicators. In
the automation industry, in order to perform these two different control tasks a
PLC with its own progmm inside and a SCADA system are used. In this
context, the humanoid robot Rh-1 is provided with the software system allowing
the implementation of the industry control concepts with the humanoid robot.
The Software architecture is based on the Server-Client model (Figure 3).

Fig. 3. System Architecture.

For the security reason, the Control Server accepts the connections of other
clients, such as the Head Client, responsible for the human-robot interaction,
only if the Master Client allows it. If the connection is accepted, the Master
Client only supervises the humanoid robot state and data transmission between
the robot and other Client, but always in the case of any conflict it has the major
priority.
 739

According to Server-Client model, the humanoid robot is controlled by the

passive Server, which waits for requests and upon its receipt, processes them
and then serves replies for the Client. But on the other hand, the Server controls
all Control Agents which reside on the CAN bus network. In that case, the
Control Server is no longer a slave. It is a network master for Control Agents
which performs their operations (motion control or sensing) and reply for the
Server.
When the robot is working in the operational interaction mode (walking,
objects manipulation, etc.), there are several computing and communication
tasks that need to be performed in a cyclic mode and fast enough to avoid any
possible loss of the control. The periodic (with period T s ) chain (Figure 4)
begins with the sensing task, taking the time interval tatt for attitude estimation
gyros and accelerometers readings and tZmpfor Z M P force-torque sensors
readings.

Figure 4 Main computing and communication tasks for the Rh-1

These tasks are followed by the tasks performing CAN bus communications,
Posture Control and then Stabilizer and Inverse Kinematics Calculator
computing, internal PC bus communications, Supervisory controller, and CAN
bus transmission of new reference for each joint of the humanoid robot.
The period TS should be small and compatible with the dynamics of the
humanoid robot movement. On the one hand, TS cannot be made arbitrarily
small because various computing and communication tasks with execution times
tan, fzmp, tCAN, tpc, tstabilizer, etc., cannot be made arbitrarily small. Also, a small
value for TS would generate too many messages in communication lines (RS-
232, PC bus, CAN bus) that would overload it. On other hand, TS cannot be
made arbitrarily large because of the dynamics of the robot (Nyquist criterion).
Thus,

Ts 'tatt + t . z ~ p+ tC,4N + tPControl + tStabilizer +

And
740

1
TS< -
2.Fr
Where Fr is the highest movement frequency on any robot link.
Assuming that the robot walks about the same speed of a normally walking
human, Fr = 2 Hz.
Thus,

Ts < 250 ms (3)

A reasonable value for the sum of the time intervals taken by of all
computations and calculations is 0.75 Ts , thus

tatt + tzmp + tCAN + tPControl -k tStabilizer +

(4)
+ tlnvKjn+ tPc + tsUp+ tCAN= 187.5 ms
This upper limit of the sample time is a strict real-time requirement considering
the complexity of the computing and communication tasks to be performed with
this time limit.
As a PLC in the automation industry, the Control Server is designed and
programmed as finite state automata with a sample time TS. Figure 5 shows the
state diagram of the humanoid robot Control Server functioning.

tas I
t
_t__J
Figure 5 Server functioning state diagram.
 741

Two basic types of the incoming data are processed. A command is the simple
data, which can be executed by one Control Agent. The order is a complex
command which needs the simultaneous action of many Control Agents and
sensors which possesses the humanoid robot. After the connection of the Master
Client, the humanoid robot stays at the Client Handling state waiting for an
order or a simple command. The arrival of an order launches the User Program.
The User Program is the core of the humanoid robot Control Server software. It
performs the trajectory execution at the synchronized multi-axis walking
applications, controls the posture and ZMP errors at the dynamic walking mode,
reads the sensors state etc. Developed software provides the set of the C-based
function to work with the robot and to generate user’s motions and control
procedures not only for walking, but for implementing different human-robot
cooperation tasks.
To provide the robot Rh-1 with bottom level control, a SCADA system for
the humanoid robot, titled HRoSCoS (Humanoid Robot Supervisory Control
System) Client was developed.
The HRoSCoS Client provides the trending of different parameters of the
robot, such as for example the joints velocities, accelerations, currents, body
inclinations, forces and torques appears during humanoid robot walking. Real-
time and historical trending is possible. Logged data is time-stamped and can be
filtered when viewed by a user. Also it is possible to generate different reports
on the humanoid robot state at any time.
The HRoSCoS Client system presents the information to the operating
personnel graphically. This means that the operator can see a representation of
the humanoid robot being controlled (Figure 6 ) .

Fig. 6. HRoSCoS Client views.

742

The HMI supports multiple screens, which can contain combinations of

synoptic diagrams and text. The whole humanoid robot is decomposed in
"atomic" parameters (e.g. a battery current, its maximum value, its odoff status,
etc.) to which a Tag-name is associated. The Tag-names used to link graphical
objects to devices. Standard windows editing facilities are provided: zooming,
re-sizing, scrolling, etc. On-line configuration and customisation of the HMI is
possible for users with the appropriate privileges. Links are created between
display pages to navigate from one view to another.

4. Communication Infrastructure
When building a control application, communication with the host is often a
crucial part of the project. Nodes of the network always function as data servers
because their primary role is to report information (status, acquired data,
analyzed data, etc.) to the host at constant rates.
Hardware Architecture consists of three basic levels of automation which uses
its own communication systems. The upper (Control) level uses a TCP/IP based
communication protocol. Ethernet communication is one of the most common
methods for sending data between computers. The TCP/IP protocol provides the
technology for data sharing, but only the specific application implements the
logic that optimizes performance and makes sense of the data exchange process.
When data transmission begins, the sender should packetize each piece of data
with an ID code that the receiver can use to look up the decoding information.
In that way, developed communication protocol hides the TCP implementation
details and minimizes network traffic by sending data packages only when it is
needed. When a data variable is transmitted by the sender, it is packetized with
additional information so it can be received and decoded correctly on the
receiving side. Before each data variable is transmitted, a packet is created that
includes fields for the Data Size, the Data ID and the data itself. Figure 7 shows
the packet format.

~
Data
- c32 byrsii)
Fig. 7. The package format.
. -- - -7
--

The Data ID field is populated with the index of the data array element
corresponding to the specified variable. Since the receiving side also has a copy
of the data array, it can index it to get the properties (name and type) of the
incoming data package. This very effective mechanism is implemented to
provide data exchange between the Control Server and different Clients on the
Control level of automation of the humanoid robot.
 743

Bottom level (Sensorial and Field) communications are realized using CAN
and CanOpen protocols. These communication protocols provide data
transmission in broadcast type of communication. This guarantees data
integrity as all devices in the system use the same information. The sensorial
system of the humanoid robot makes data exchange under the lower CAN
protocol and the intelligent motion controllers use the upper CANOpen
protocol.
The communication implemented on the bottom level involves the
integrating of CANopen (Drives and Motion Control Device Profile) and
the introduction of new functionality which is not contained within the relevant
device profiles for the sensorial data processing.

5. Conclusions
This paper presented the development of the control system of the humanoid
robot Rh-1. The control system was designed using software development
standards from the industrial automation field.
It was proved experimentally, that Rh-1 robot provided with the current
mechanical, hardware and software control architectures can walk stably.
This work makes an effort to show that it is possible to bring some basic
aspects of the industrial automation programming to other, more sophisticated
fields of robotics in order to extend the further standardization and unification
of the design processes. Moreover, proposed approach allows consider a
humanoid robot locomotion inside the global control problem.
In the future a human - humanoid robot interaction from the software control
system design side will be considered. Also, the analysis and improvements of
the presented software control architecture will be continued.

References
1. K. Kaneko, F. Kanehiro, S. Ka’ita, K. Yokoyama, K. Akachi, T.
Kawasaki, S. Ota and T. Isozumi, ‘desi of prototy e humanoid robotics
latform for HRY, Proc. of IEEE/R& Int. Conj5‘rence on Intelligent
gobots and Systems, pp. 243 1-2436, (2002
B
2. Y. Sakagami, R. Watanabe, C. Ao ama, . Matsunaga, N. Hikagi and K.
Fujimura, “The intelli ent ASIdO: s stem overview and integration”,
Proc. of IEEE/RSJ Int. Eonference on Xtelligent Robots and Systems, pp.
2478-2483, (2002).
3. I. Park, J. f i m , S. Park, J. Oh, “Development of humanoid robot latform
KHR-2”, Proc. of IEEE Int. Conference on Humanoid Robots, pp. $92-3 10,
4. n
ov;):!?g. M.A.Rodriguez; M.Arbul6; PStaroverov; L.M.Cabas;
C.Baiguer. “Advanced motion control s stem for the humanoid robot Rh-
o”, 8th International Conference on C l i m h g and WalkingRobots, (2005).
SPARBOT - A ROBOTIC FOCUS MITT TRAINING PLATFORM

RICHARD STOKES, LIQIONG TANG and IBRAHIM AL-BAHADLY

Institute of Technology and Engineering
Massey University
New Zealand

This paper describes the design and implementation of a prototype robotic training
platform for focus mitt training called 'SparBot'. Focus mitt training is an important
aspect of the Martial Arts. The goal of SparBot is to emulate the focus mitt pad-holding
trainer and introduce a level of automation to focus mitt training. The system has two
movable arms with rotating wrists, which makes SparBot able to move its arms into
different positions to receive the strikes. Sensors throughout SparBot monitor arm
position and robot location on the linear translation drive. The robot is capable of giving
audio cues for strike patterns and can track training progress in the control software.
SparBot seamlessly integrates modem robotic technologieswith Martial Art training.

1. Introduction

Intelligent robotic and mechatronic systems form the core of industrial and
consumer products in the 21" century. The use of robotic apparatus in the
domestic consumer sector is on the rise. The core disciplines of Robotics and
Mechatronics allows engineers to quickly turn an idea into a physical prototype.
The integration of robotics, precision mechanical systems, intelligent control,
and reliable electrical and electronic systems has made many previously
unattainable ideas a commercial reality. The work presented in this paper
describes the development of a prototype robotic Martial Arts training system -
SparBot. It is the first prototype of a Robot Martial Arts Trainer for focus mitt
training.
Martial Arts Training is part of the training course for army soldiers and
police officers in many countries. It brings benefit to both physical and
emotional health. Martial Arts Training improves self-defense capability and
foster self-confidence that allows people to diffuse volatile situations without
resorting to force. The strict mental and physical training schemes also
promotes cardiovascular and mental health, which enable people cope better
with the high stress level in today's competitive society. Nowadays, Martial
Arts Training courses become popular especially for young people. The
objective of the project is to develop a prototype robotic trainer for focus mitt
training and domestic sports equipment for daily exercise.
Different training schemes, tools, and equipment are used in Martial Arts
Training. Focus mitts are a type of training pad. Martial Arts trainers fit the0
pad over their hands during focus mitt training. By calling out strikes and

744
 745

placing the pads in certain designated positions, the trainer can control how their
partner moves and strikes. The presented prototype robotic trainer is designed
to emulate the pad-holding partner for focus mitt training. The robot is able to
position its arms to receive strikes, has some degree of lateral movement and can
provide audio cues for the user. The entire robot is controlled via a graphical
user interface running on a local PC.

2. Robot ~ e c ~ a t r o nSystem
ic Design
The main strikes in focus mitt training are the jab, cross and hook. To mimic the
pad-holding partner in these strikes, the robot trainer must have a mechanism
which can realise wrist and arm movement as well as the ability to change its
location around. In addition the system is able to give commands to conduct the
training. Thus, the final machine must have the following fundamental
components and features to be a basic robot focus mitt trainer.
e Two movable arms, each with rotating wrists
e Translational movement of the robot body
e The ability to give audio cues
e Auto and manual mode
During focus mitt training, the pad-holding trainer must hold his arm in space
at a right position to withstand the strong impact forces [I]. In order to
maximise the area moment of inertia to provide a strong mechanism for impact
force opposition, the robot arms are made from aluminium tubing. This also
facilitates placement of the wrist stepper motor, cable routing and coupling
inserts.
The structure of the robot arm is designed in the form of a human being’s
arm. The robot arm assembly as shown in Figure 1 has a wrist and a shoulder
insert. ,.

8 .

_ I

Fig 1: Robot Arm Assembly Fig 2: Bearing assembly

A rotating pad plate is fixed in the wrist via a bearing housing shown in
Figure 2. Such a design provides the rotating movement for the pad through a
stepper motor inside the wrist and transfers strike forces away from the rotor
746

shaft into the arm housing. The shoulder insert connects the arm assembly to the
shoulder drive shaft inside the upper torso and also transfers the impact forces to
the robot base [2].
The pad drive stepper motor is coupled to the pad base plate and mounted
inside the forearm. On top of the pad base plate is the pad assembly as
illustrated in Figure 3. Providing a comfortable strike surface, the pad backing
plate is fitted with composite foam, behind which is embedded a micro switch
that provides a means for registering that a strike has occurred. If an off-centre
strike is apparent at the pad surface, the stepper holding torque is not sufficient
to withstand such an impact. Thus, a solenoid operated locking pin is
implemented, The system engages when the pad is in place locking the rotating
plate to the base plate on top of the bearing housing. The stepper motor
facilitates accurate pad positioning for the jab, cross and hook punches.
Positional accuracy is of the utmost importance to ensure correct locking pin
alignment.

Figure 3. Robot Pad Assembly

Arm rotation is facilitated by two 24V DC motors coupled to the shoulder

drive shaft via a 2.5:l chain-driven gear arrangement. To remove strike forces
from the DC drive gearbox and drive train, a braking system is employed. The
braking system is operated by solenoid through a pin to lock the brake drum and
hold the arm at the correct positions as shown in Figure 4.
 747

Figure 4. Robot Arm Brake System

The ideal robot trainer should be able to change its location and move around,
which is a two-dimensional movement. As this is the first stage of the project, an
existing linear translation drive system is used to provide another dimension to
SparBot's movement. The intention is to emulate the linear movement of a pad-
holding partner. The system has a three phase 0.18 kW drive motor and a power
screw. After installed some sensors and limit switches, it was nicely integrated
with the robot body and realise the linear movement.

3. Electrical System
The robot trainer electrical system consists of two main parts, the Low Voltage
(LV) subsystem and the Functional Extra Low Voltage (FELV) subsystem. The
LV subsystem contains the three phase and associated single phase power
systems. The LV scheme is shown in Figure 5. The incoming three-phase supply
is run through a local isolator then passed to the live side of both the forward
and reverse contactors. The motor supply is run through a three-phase overload
relay then to the motor. A red phase tap supplies, via a local MCB, two solid
state relays (SSR) that are switched via the DIO of the Labview card. This
provides the electrical isolation necessary for FELV classification. The
switching current to each contactor coil is interlocked via an auxiliary contact
mounted on the opposite contactor to prevent accidental engagement of both
contactors simultaneously, which would result in a phase-to-phase fault.
748

Figure 5. The Low Voltage Subsystem

The FELV subsystem contains the three DC supply rails of 5,6 and 24 VDC
and the DIO signals from the Labview control card. The DC rails facilitate the
following functions:
* 5V rail: logic supply
0 6V rail: supply for DC solenoids and stepper motors
* 24V rail: supply for the DC servos
The control card for SparBot contains drivers for the DC servos stepper
motors as well as MOSFET's for solenoid control. The physical connections
from the control card to the board and out to the field are facilitated by standard
DIN rail terminals with blade disconnects. This allows one to connect each
subsystem in turn during commissioning, allowing for easier debugging, as
shown in Figure 6.

Figure 6. Electrical Layout

4. Robot Control System

The robot trainer control system is developed using LabView via the PCI-6229
multi-functional card and associated breakout board. The card provides 48
digital I/O lines and several analog IIO. The software developed allows SparBot
to run under two modes: manual and automatic.
 749

Under manual mode the SparBot Test Centre is invoked. Once the Test
Centre is activated the software controls on the user interface are linked to each
field device or control chip via the SCB-68 breakout box. Each field device may
be manually driven, and the status of each sensor is continually polled. Figure 7
is the user interface of the Test Centre.

Figure 7. SparBot Test Centre User Interface

From the Test Centre, the user can perform linked functions, such as moving
the arm to a desired position. From the click of a button, all necessary control
actions are p e r f o ~ e din sequence. This allows the function and interaction of
each hardware and software subsystem to be fully tested and tuned. For instance,
to move from the "left jab" position (arm up, pad forward) to the "left hook"
position, the following actions are required

0 Check current position, play audio cue.

0 Disengage pad lock, set stepper direction, enable stepper for turn time, t.
0 Disengage arm lock, set DV drive direction, enable DC drive.
0 When t has elapsed, disable stepper and engage pad lock
0 When "am down" sensor reports true, de-energise DC drive, enable
brake.
When the system is used under the automatic mode, after a short audio
welcome note, the user is presented with a keypad as illustrated in Figure 8 into
which the user enters the PIN number. This allows easy identification and
tracking of each individua~user. The PIN is compared with a lookup table and
750

if the user has a data file it is loaded with the last session details, and the
software automatically configures itself for the correct level. If no record exists,
one is created, and the user is supplied with a PIN.

Figure 8. Keypad Screen

Once the set-up phase is completed, the user may select a round time and
level of difficulty (if they do not want the defaults), then the training will begin.
This consists of concatenations of the basic movement building blocks discussed
in the 'manual mode' section. These subroutines are called as needed. For
instance, when performing a 'double-punch, hook-uppercut' manoeuvre, the
arms are placed for the jab-cross combination then once the last cross punch has
been registered, the arms automatically move to the hook-uppercut positions.
Only until the uppercut punch is registered, can the next sequence begin. During
training, certain parameters such as punch timing, accuracy, number of punches,
etc are recorded and appended to a database file. This information associated
with the user via the PIN is stored in the database and can be retrieved at a later
time.
By using a text-to-speech (TTS) program [3], any typed text can be created
from a selection of voice types and stored on the local disk as a PCM wave file.
A wave file player was developed in LabView that takes a single filename
parameter pointing to the file location on disk. Once the control movements are
complete, the wave player is invoked, informing the user of the strike +ype and
order required. The termination of the wave player sub-VI indicates the end of
the movement, and control is returned to the main loop.
 75 1

5. Conclusion
A prototype robotic training platform for focus mitt training has been developed.
The system can demonstrate the major functions used in focus mitt training.
Different training programs can be stored in the system through the software.
The system is also able to store personal training data and track the training
history. However, the prototype does reveal some problems. Further
improvements will focus on mobility, safety, and wireless control.

6. References
1. John D. Pierce Jr., Kirk A. Reinbold, Barry C. Lyngard, Robert J. Goldman,
and Christopher M. Pastore, "Direct Measurement of Punch Force During
Six Professional Boxing Matches," Journal of QuantitativeAnalysis in
Sports: Vol. 2: No. 2, (2006). Article available at:
httu://www.be~ress.comlia~/vol2/iss2/3.

2. The Engineerng Toolbox, "Steel modulus of rigidity," Z%eEngineering

Toolbox (2006), httu://~~.enpinqerinetoolbox.co1l7/modulus-ripiditv- d 9~6.blInl.

3. Ultra HAL Text-to-Speech Reader version 1.O, 10th September, 2006,

www.zabaware.com, Zabaware, Inc.
 ACTUATORS AND ORTHOSES TO ASSIST FES-CYCLING

R. MASSOUD, M. 0. TOKHI, M. S . ALAM, M. S . HUQ

Department of Automatic Control and Systems Engineering, University of Shefield, UK

In this paper, mechanical springs are used to assist quadriceps-FES-cycling, where only
the knee extensors are electrically stimulated to drive a stationary bicycle. Three types of
mechanical springs are tested: torsion spring, linear extension spring, and constant-force
spring. Utilizing torsion spring does not improve FES-cycling performance, while FES-
cycling performance shows significant improvement when other spring types are used.
This study shows that the improvement of FES-cycling performance depends on the type,
position, and parameters of the spring. Different spring situations were tested and the
performance was evaluated in terms of minimizing the error in crank cadence and
maximizing the efficiency. It is noted that selection spring’s situation is not easy, as this
is a multi objective optimization problem. Non-dominated solutions were obtained, and
the decision was taken for the solutions for which the error and efficiency were within
acceptable ranges.

1. Introduction
FES-cycling is a rehabilitation cycling exercise for people with spinal cord
injury (SCI), where functional electrical stimulation (FES) is applied to their
still functioning leg muscles to generate functional movements and enable them
drive the rehabilitation bicycle. FES-cycling has significant impact on an SCI
person’s physical and psychological health [I]. It has proven that FES-cycling
performance can be enhanced when auxiliary motors are utilized. Therefore, by
using these motors, the overall power can be greatly increased, loss of power
due to muscle weakness or fatigue can be compensated for, and leg cycling
motion can be maintained even at very low leg power levels [2]. People with
very weak muscles at the beginning of their training period will be able to
exercise FES-cycling assisted by an auxiliary motor [3].
Pons et al. [4] indicated that it is possible to have the motor assist the
pedalling during part of the cycle and retard it during another part of the same
cycle. The motor resistance can be replaced with an energy storage device such
as mechanical spring, which releases its energy when assistance is needed. In
his ergonomic study of cycling for normal people, Rasmussen et al. [ 5 ] tried to
eliminate the dead centres of the pedal cycle by providing the bicycle frame
with a spring arrangement, They used two springs attached to the bicycle frame,
which allow to store elastic energy that help to overcome the dead centre of the

752
 753

pedal cycle. The springs’ stiffness, slack length, fixation points on the frame,
and fixation points on the crank were optimized. This lead to uniform effort
over the cycle period, allowing the rider to produce an even crank torque.
Furthermore, the maximum muscle activity was less than half of the case with
no springs mounted [5].
To the authors’ knowledge, the use of mechanical spring to assist FES-
cycling has not been reported in the literature. This paper discusses the
possibility to replace conventional auxiliary electric motor with an energy
storage actuator, so as to increase the FES-cycling efficiency and obtain even
FES-cycling. Mechanical springs may come in a wide variety of different types,
such as helical compression or tension springs, helical torsion springs, leaf
springs and flat springs. Generally, mechanical springs have many shapes,
different specifications, and various sizes. Moreover, the can be made from a
wide range of materials. This allows to produce springs with mechanical
properties very similar to that in human muscles, to produce similar forces
produced by human muscles when released after being distorted.
Gharooni et al. [6] used a spring to mimic the knee flexors (hamstrings
muscle group) during paraplegic gait restoration. They extended the knee joint
by stimulating the quadriceps muscle group, and used the spring for knee joint
flexion.

2. Simulation set up

A humanoid and bicycle model is designed via Visual Nastran software. The
humanoid dimensions are described in [7]. The bicycle consists of a fixed seat
and a crank with constant load resistance of 0.18 Nm. The distance between the
centres of the humanoid trunk and the crank is 0.7m, and the seat and the crank
are at the same height (see Figure 1).
In order to simulate FES, Riener’s muscle model is used [8]. The model is
for quadriceps muscle group, and it depends on the activation and contraction
dynamics of the paralyzed muscle. The bicycle model is driven by the knee and
hip joints torques, produced by the muscle model. The muscle model produces
the torques according to the level of the FES, which is controlled by a
proportional-derivative (PD) type fuzzy logic (FL) controller. The controller’s
design and the FL control strategy are described in [7].
Since only the quadriceps (knee extensors) is stimulated in this work, the
springs are used to mimic the knee flexors (hamstrings). Each spring has two
work shifts; energy storage and energy release. During energy storage shift the
spring stores the exceeded energy produced by quadriceps stimulation and the
754

lower limbs dynamics, while in the energy release shift the FES is off for both
legs allowing the spring to flex the knee joint and drive the bicycle.
Accordingly, FES pedalling cycle is divided into three essential phases: the left
quadriceps stimulation phase, the right quadriceps stimulation phase, and the
relax phase where no quadriceps are stimulated allowing the springs to release
their potential energy, and to take part in the driving process.
3. FES cycling with torsion springs
Two linear torsion springs have been mounted very close to the knee joints of
the humanoid as shown in Figure 1. In pratice, this can be done by following the
same arrangement as implemented by Gharooni et al. [6].

Figure 1: Humanoid with bicycle designed in visual Nastran software.

Figure 2 shows that the best FES-cycling performance is obtained when no

torsion spring is used, error in crank cadence increases as the spring constant
increases. Conversely the gross efficiency (the ratio between the output power
and the muscle power) decreases with the increment of the spring constant. This
can be seen clearly in Figure 3. Due to these results, it is clear that using torsion
springs positioned at the knee joint does not enhance the performance of FES-
cycling, in terms of efficiency and percent error.
 755

0.26
0.25
L 0.24

-;
10.23
0.22
0.21
k 0.20
0.19
0.18
0.17
0.16
0.00 0.01 0.02 0.04 0.06
Sprlng constant [Nmldeg]

Figure 2: Percent error for torsion suring imulementation

Figure 3: Relation between efficiency and torsion spring’s constant

4. FES cycling with linear extension springs

In this case, the torsion springs are replaced by two linear extension springs
mounted on the frame of the bicycle. The spring’s one end (say, first end) is
fixed to the pedal and the other (say, second) end is fixed to the bicycle frame in
a semi vertical position. Figure 4 shows the spring position according to the
bicycle. The vertical distance between the centre of the crank and the second
end of the spring (VDCS) is 0.29m. This vertical distance is still constant while
the horizontal distance between the centre of the crank and the second end of the
spring (HDCS) is changed from Om (vertical position) to 0.25m to the right, and
from Om to O.lm to the left by a step of 0.05m. Eight spring positions were
considered. At each spring position, the spring constant was changed from 0
N/m (no spring used) to 50 N/m, at steps of 5 N/m. At the spring positions (-
0.2, 0.29) and (-0.25, 0.29) the spring constant was increased to 60 N/m.
Accordingly 92 different situations for the spring were tested.
FES-cycling performance was evaluated in terms of efficiency and error in
crank cadence. Figure 5 shows that the percent error in crank cadence decreases
with increase in the spring constant until it reaches a point which has the lowest
error value. After this point the error starts to increase with increase in the
spring constant. Moreover, the error decreases as the spring leans to the right
(towards the humanoid).
756

Figure 6 shows the FES-cycling efficiency. It is noted that the efficiency

increases when the spring constant increases for all spring positions; maximum
efficiency is reached when the spring is in its vertical position and it decreases
when the spring inclines to either direction (left or right).

0.1 0 -0.1 -0.2

Figure 4: Linear spring implementation

with constant force extension springs

The constant-force extension spring is basically a high stress, long deflection
device and offers great advantages such as very low or zero gradient, large
extension, small package size. Two force-constant springs were fixed to the
bicycle; the first end of the spring was fixed to the pedal while the second was
attached to the bicycle frame. Figure 7 shows that the horizontal distance
between the centre of the crank and the second end of the spring (HDCS) is
constant (OSm), while the vertical distance is changed up and down from -0. Im
to 0.2m in steps of 0.05m.

0.3

5 0.25
C
I. 0.2
E 0.15

0 5 10 15 20 25 30 35 40 45 50 55 60
Spring constant “lml
I
Figure 5: Relation between percent error and linear spring constant at different HDCS
 757

0.14

0.135
-4.2 rn
0.13

E, 0.125
~..-4.05rn
d 0.i2
E 0.115
-+0.1 m
0.11
- - -4.25 m
0.105
0.1

L
0 5 10 15 20 25 30 35 40 45 50 55 60
spring Conelant purn]
- . _ _ ~

Figure 6 : Relation between FES-cycling efficiency and linear spring constant at different HDCS

At each spring position, the spring constant was changed from ON (where
no spring was used) to 25N in steps of 5N. Therefore, 42 different possibilities
for the spring position and parameter were tested. The FES-cycling performance
was also evaluated in terms of efficiency and percent error in crank cadence.
Similar to the case of using linear spring in FES-cycling, the percent error
decreases when the spring constant increases until it reaches the turning point,
which has the lowest percent error. Beyond this point the error increases with an
increase in the spring constant, as shown in Figure 8. Gross efficiency obtained
at different situations of the spring is shown in Figure 9. It is noted that the
relation between the efficiency and the spring constant is nearly linear.

Figure 7: Constant-forcespring implementation

6. O p t i ~ ~ a t i oofnthe spring para~etersand fix points

The tested spring parameters and positions were chosen heuristically. And the
results obtained from applying theses parameters show the difficulty in selecting
the optimum spring situation. In order to find the best solution, the efficiency
should be as high as possible while the percent error should be as low as
possible. Since both objectives are in conflict with each other, the optimization
758

problem can be converted to a two-objective minimization problem by

minimizing both, llefficency (maximizing the efficiency) and the percent error
in crank cadence. This can be represented by a two -objectives space, where the
horizontal axis represents the percent error and the vertical axis represents
llefficiency. Figure 10 shows the results for all the spring situations for both
linear and constant-force springs in the two -objectives space. Solutions which
are very close to the origin are preferred, since they dominate the others. From
the non-dominated solutions of both spring types (see Figure lo), solutions with
percent error less than 0.13 and with gross efficiency more than 12% are
chosen. Three points are satisfied; two of them represent two linear spring cases
where the spring’s HDCS= 0.2m and VDCS- 0.29m and the spring rate for the
first case is 55Nlm, and for the second case is 60Nlm. The third point represents
the constant force spring where the HDCS= 0.4m and the VDCS= 0.2m and the
spring constant is 15N.

1 0.40 , 7
-0.15

-0.05

-4.05
-4.1
I l l 1
-0.2

0 5 10 15 20 25
Spring constant [N]

Figure 8: The relation between percent error and constant -force spring constant at HDCS=O.Sm and
at different VDCS

-0.15

-++ -0.05
-0- -0.1
+0.2
0 5 10 15 20 25
Spring constant [N]

Figure 9: The relation between the efficiency and the constant-force spring constant at HDCS=O.Sm
and at different VDCS
 759

7. Conclusion
This study concludes that using torsion springs mounted near the knee joints
does not enhance FES-cycling, while using linear or constant force extension
springs mounted on the bicycle fiame can increase the efficiency and reduce the
percent error in crank cadence. Therefore, it has significant impact on the
overall performance. It is clear fiom this study that optimizing spring position
and parameters leads to better FES-cycling performance.

Figure 10: Spring situations for linear and constant -force springs; non-dominated solutions are
marked with filled points.

References
1. Petrofsky, J. S., Laymon, M. (2004). The effect of previous weight training
and concurrent weight training on endurance for functional electrical
stimulation cycle ergometry. European Journal Applied Physiology, 91 (4),
392-8.
2. Hunt, K. J., Stone, B., Negard, N. O., Schauer, T., Fraser, M. H., Cathcart,
A. J., Ferrario, C., Ward, S. A. and Grant, S. (2004). Control strategies for
integration of electric motor assist and functional electrical stimulation in
paraplegic cycling: utility for exercise testing and mobile cycling. IEEE
Transactions on Neural Systems and Rehabilitation Engineering, 12 (l),
89-101.
3. Gfohler, M., Loicht, M., Lugner, P. (1998). Exercise tricycle for
paraplegics. Medical & Biological Engineering & Computing, 36, 118-121.
4. Pons, D. J., Vaughan, C. L. and Jaros, G. G. (1989). Cycling device
powered by the electrically stimulated muscles of paraplegics. Med Biol
Eng Comput, 27, (l), pp. 1-7.
5 . Rasmussen, J., Christensen, S. T., Damsgaard, M., Zee, M. (2005).
Ergonomic optimization of a spring-loaded bicycle crank. 6Ih World
Congresses of Structural and Multidisciplinary Optimization, Rio de
Janeiro, 30May-03June, Brazil.
760

6. Gharooni, S., Tokhi, M.O. (2002) Synthesising the swing phase in orthotic
gait using a hybrid FES spring brake orthoses. PhD thesis, University of
Sheffield, Sheffield.
7. Massoud, R., Alam, M. S., Tokhi, M. O., Gharooni, S. C. (2006). GA
optimized spring orthosis to enhance FES-cycling performance. CD
Proceedings of the 91h International Conference on Climbing and Walking
Robots (CLAWAR2006),Brussels, Belgium, pp 334-337.
8. Riener R., and Fuhr T. (1998), Patient-driven control of FES-supported
standing up: A simulation study, IEEE Transactions on Rehabilitation
Engineering, 6(2), 113-124.
A NOVEL MINIATURE ATTITUDE MEASUREMENT SYSTEM
FOR CLIMBING AND WALKING ROBOTS

GUANGLONG WANG, CHUNXI ZHANG

School of Instrument Science and Opto-electronics Engineering,
Beijing University of Aeronautics and Astronautics, Beijing 100083, China

ZHAOYING ZHOU, RONG ZHU

MEMS Lab, Department of Precision Instrument,
Tsinghua University, Beijing 100084, China
A novel design of miniature attitude measurement system for climbing and
walking robots is presented. By integrating three-dimensional magnetometers
and three microelectromechanical accelerometers, the system provides high-
frequency position and attitude data. And it takes advantages of the solid-state
configuration, small size, light weight, high reliability, low power consumption,
rapid startup and relatively low cost. It can be used to test the transient attitude
of the moving body such as robots. The three-axis accelerometers detect the
three orthogonal components of the gravity, by which tilt of the robots including
pitch and roll can be derived. Otherwise a three-axis magnetometer which are
also assembled as an orthogonal placement to measure the three components of
the Earth's magnetic field. The calculation and simulations indicate that the
formalism proposed in this paper is able to reach an excellent long-term
performance. Even if there is various error sources existing in the measurement,
the non-orthogonal assembling error is the most dominating. We analyze the
representative features of this error, establish a mathematic model for it, and
successfully compensate this error via arithmetic. Finally, real experiments
performed with a mobile robot are presented and analyzed. The results of the
test indicate that the system can measure the freewill attitude in three-
dimensional space. It is expected that this system will be adopted in many fields.

1. Introduction
In the past decade, precise attitude measurement performed with mechanical
accelerometers and gyroscopes based on solid-state sensor technology have
become readily available and have supplanted mechanical sensors in some
climbing and walking robots applications. Unlike mechanical devices, these
microelectromechanical solid-state sensors do not rely on numerous and large
moving parts. Instead, they use miniature sensing elements made from quartz

76 1
762

crystals or micro-machined silicon. Knowing the initial orientation and position

of the robot, the mathematical integration of the acceleration and rotation rate
measurements yields the attitude and the trajectory of the robot. Such inertial
techniques are completely autonomous and need no external reference [ 11. But
they will induce the inevitable long-term drift by dead reckoning, and have
relatively slow start-up for its initial alignment. This paper provides a novel
miniature attitude measurement system, which is composed of micro-
electromechanical accelerometers and magnetometers. We use a kind of
absolute measurement to abandon the problems of the conventional inertial
measurement mentioned above. Micromachining is one of the most important
emerging technologies for sensors because it provides unique advantages due to
offering small size and lowing cost [I] [2]. The main benefits that silicon based
microelectromechanical system derive from its common base with the IC
industry are "low-cost", "massproduction", and "monolithic integration". The
important requirements for micromachining technology is that it should be
simple, easily accessible to a wide range of users, be low cost and provide
sufficient flexibility so that different types of sensors can be fabricated on the
same chip without significant redesign[3]. During the past few years, two
generic technologies have become mainstream in the field of micro
electromechanical systems. These technologies are commonly known as bulk
micromachining and surface micromachining. Among them, the surface
micromachining technology attracts more attention by its simplicity, high
reliability, relatively low cost and high finished product probability. As result of
this, we adopt a dual-axis sensor and a single-axis sensor, which take benefit of
silicon surface micromachining technology[4] [ 5 ] .

2. The Composing of the System

The attitude measurement system is made up of a suit three-dimensional
accelerometers, a suit three-dimensional magnetometers, A/D converter and
microprocessor etc. The analog signals readings from the robot attitude sensors
are converted to the digital signals through the A/D converter. Then they are
processed in real-time by microprocessor to get the three attitude angles. Finally
the attitude angles are conveyed through serial port into control unit to display
or drive servos actuators. Figure 1. shows the system composing of the system.
The three-dimensional orthogonal accelerometers measure the components of
acceleration of gravity, which deduce the tilt. We select ADXL202 which has
dual-axis and surface micromachined polysilicon structure as the acceleration
sensor because it has the advantages of hlly integrated monolithic, low cost,
low power and faster response. And the three-dimensional orthogonal
magnetometers measure the components of the earth's magnetic field, by which
the robot attitude is computed. For magnetic sensors, a common type of it for
 763

the robot systems is the fluxgate sensor. It provides a low cost means of
magnetic field detection, but it also tends to be bulky, somewhat fragile, and has
a slow response time. Another type of magnetic sensor is the magnetoresistive
sensor [4]. This sensor is made up of thin strips of permalloy whose electrical
resistance vanes with a change in applied magnetic field. The magnetic product
we prefer is HMClOOl and HMC1002. They are single and dual-axis
magnetoresistive microcircuits, provide extremely sensitive magnetic sense in
small package outlines. These sensors offer low cost and small size, and high
sensitivity.

accelerometers + Signal + A/D

+ proces + converter
- sing
magnetometers Serial
- Micro Port

3. Principle of the Attitude Measurement

The earth's gravity field and magnetic field in geographic coordinate system
(GCS) is shown in Figure 2. Therein to j? is the dip angle of the magnetic field.
The three axes of the magnetometers are mounted along the X, Y and Z. And
the magnetic readings x, , y mand z, which figure the X, Y and Z components
of the earth's magnetic field, can be transformed back to the geographlc
coordinates by applying the Cosine Matrix C," shown below:

sin ycose - sin 0

1
C: = cosy sin esiny - sinycosy sin w sin 0 sin y + cos w osesiony
Lcos y sin ecos y + sin I+J sin A siny sin ecosy - cosysiny cos e cos y
emrepresents the disturbance and error in magneto measuring.
The three axes of accelerometers are also along the X, Y and Z to measure the
components xg,ygand zg of the earth's gravity field. The earth's gravity field& in
764

GCS has very simple pattern, and it is always pointing to downword direction
and the horizontal components are zero. Then

where eg represents the disturbance and error in gravity measuring.

Putting into thec,” and eliminating the disturbance and error by signal
processing, the attitude angles can be derived as:
Y = a%(yg/zg 1
e = -arctg(cosy. x g / z g )
w = arctg(yh/xh)
where xh = x, cos 8 + ( y , sin y + z, cosy) sin 8 ,
y h = z, sin y - y, cosy

Crnvity tietd

5 Dawn
Figure 2. The earth‘s gravity field and magnetic field in GCS

4. Error Analysis
There are several of errors existing in system, including random noise and
measured deviation. Filter technology has been applied to restrain the random
noise. And for measured deviation, it mainly consists of the non-orthogonal
assembling errors came from the separate single-axis senor. This error exhibits a
kind of regular characteristics and is distributed in the attitude measuring. The
most straightforward scheme for eliminating this error is avoiding non-
orthogonal, as adopting a three-axis monolithic sensor. But a three-axis sensor
has more complex by surface micromaching technology. The cost is
unavoidably highly raised, but the performance is restricted. Take account of
this reason, we select a dual-axis sensor and a single axis sensor to compose an
orthogonal three-axis placement. For eliminating the non-orthogonal assembling
 765

error, a software compensation scheme based on error modeling is designed[6]

t71 [81.

5. Discussion and Conclusion

Theory analysis and experimental results indicate that this kind of attitude
measurement system can be used in climbing and walking robots attitude test
and control. Also the system can be packaged in a immature size and fixed in
the body of small robots. The performance of the accelerometer and
magnetometer is shown to be acceptable as a short-duration distance-measuring
device for robot platform. It can be combined with a gyroscope and odometer to
form a dead-reckoning positioning system for a robot or platform. Further
research would focus on the proper modeling of the accelerometer in order to
reduce the effect of random bias drift and error.

Acknowledgments
The authors would like to thank Dr. Rong Zhu for her help in developing some
of the theory analysis and calculation. The authors would also like to thank the
MEMS Lab, Department of Precision Instrument, Tsinghua University to
facilitate the success of the experimental work. The research work was
financially supported by a China National Science Foundation.

References
1. Bing L. Luk, David S. Cooke, Stuart Galt, Arthur A. Collie and Sheng
Chen, Robotics and Autonomous Systems, Vol. 53, Issue 2,30 November,
142-152(2005).
2. Demoz gebre-egziabher, Roger C. Hayward. Hayward, J. David Powell,
IEEE Transactions on Aerospace and Electronic Systems, Vol. 40, NO. 2
627-649( 2004).
3. Gebre-Egziabher, D.; Hayward, R.C.; Powell, J.D., IEEE Transactions on
Aerospace and Electronic Systems, Vol. 40, Issue 2, 627-649 (2004).
4. Rong Zhu, Zhaoying Zhou, Sha Li and Xuefeng Sun,
Microelectromechanical Systems Conference, 50-53 (200 1).
5. Ph. Bonnifait, G. Garcia, Control Engineering Practice 7th 383-390 (1999).
6. C. Lemaire, in Proc. IEEE ITSC’97, 1068-1072. (1997).
7. A. Nishi, Computers & Electrical Engineering, Vol. 22, Issue 2, 123-
149(1996).
8. B. Barshan and H. F. Durrant-Whyte, IEEE Trans. Robot. Automat., Vol. 11,
328-342( 1995).