Autonomous Transportation and Deployment with Aerial

Robots for Search and Rescue Missions

• • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • •

Markus Bernard
Technische Universität Berlin, Einsteinufer 17, 10587 Berlin, Germany
e-mail: markus.bernard@gmx.de
Konstantin Kondak
Institute of Robotics and Mechatronics, DLR (German Aerospace Center), Muenchner Strasse 20,
82234 Oberpfaffenhoffen-Wessling, Germany
e-mail: Konstantin.Kondak@dlr.de
Ivan Maza and Anibal Ollero∗
Robotics, Vision and Control Group, Universidad de Sevilla, 41092 Seville, Spain
e-mail: imaza@us.es, aollero@cartuja.us.es
Received 26 November 2010; accepted 20 May 2011

It is generally accepted that systems composed of multiple aerial robots with autonomous cooperation capabil-
ities can assist responders in many search and rescue (SAR) scenarios. In most of the previous research work,
the aerial robots are mainly considered as platforms for environmental sensing and have not been used to as-
sist victims. In this paper, outdoor field experiments of transportation and accurate deployment of loads with
single/multiple autonomous aerial vehicles are presented. This is a novel feature that opens the possibility to
use aerial robots to assist victims during rescue phase operations. Accuracy in the deployment location is a
critical issue in SAR scenarios in which injured people may have very limited mobility. The presented system
is composed of up to three small-size helicopters and features cooperative sensing, using several different sen-
sor types. The system supports several forms of cooperative actuation as well, ranging from the cooperative
deployment of small sensors/objects to the coupled transportation of slung loads. The complete system is de-
scribed, outlining the hardware and software framework used, as well as the approaches for modeling and
control used. Additionally, the results of several flight field experiments are presented, including a description
of the worldwide first successful autonomous load transportation experiment, using three coupled small-size
helicopters (conducted in December 2007). During these experiments strong, steady winds and wind gusts were
present. Various solutions and lessons learned from the design and operation of the system are also provided.

C 2011 Wiley Periodicals, Inc.

1. INTRODUCTION with a large geographic search region (Goodrich et al.,

2008).
Search and rescue (SAR) takes many forms, each with
Systems composed of multiple robots with au-
its own unique risks and dangers to victim and respon-
tonomous cooperation capabilities can assist responders
der (Murphy et al., 2008). Urban SAR (USAR) has been
in both USAR and WiSAR scenarios. There are related
the most studied form of SAR in the robotics research
projects dealing with multirobot teams in emergency situa-
community (Arai, Tanaka, Hirose, Kuwahara, & Tsukui,
tions such as the EMBER CMU project, which aims to assist
2008; Birk & Carpin, 2006; Gage, Murphy, Rasmussen,
first responders by providing tracking information and co-
& Minten, 2004; Micire, 2008; Murphy & Stover, 2008;
ordination abilities. In EMBER, range information is used
Poppinga, Birk, & Pathak, 2008) and has become familiar
for searching and tracking of mobile targets, using multiple
as recent natural and man-made disasters have made the
robots (Hollinger, Singh, Djugash, & Kehagias, 2009). Mul-
news the world over. On the opposite end of the spectrum,
tiagent (combined ground and air) tasking also has been
wilderness SAR (WiSAR) often involves only a few victims
demonstrated recently (Hsieh, 2007; Hsieh, Cowley, Kumar,
& Taylor, 2008).
It is noteworthy that in most of the research work,
Multimedia files may be found in the online version of this article. aerial robots are mainly considered as platforms for envi-
∗ ronmental sensing; they are not used to assist victims dur-
Also with the Center for Advanced Aerospace Technologies (CATEC),
Parque Tecnológico y Aeronáutico de Andalucı́a, C. Wilbur y Orville ing rescue phase operations. The closest experimental ap-
Wright 17-19-21, 41309 La Rinconada, Spain. proach to this concept found in the literature was the use of

Journal of Field Robotics 28(6), 914–931 (2011) C 2011 Wiley Periodicals, Inc.
View this article online at wileyonlinelibrary.com • DOI: 10.1002/rob.20401
 Bernard et al.: Autonomous Transportation and Deployment with UAVs for SAR Missions • 915

autonomous helicopters for the deployment and repair of a operations, the iMAR GmbH and the German Aerospace
wireless sensor network (Corke et al., 2004a, 2004b). Center DLR developed the “iSLD-IVC” (iMAR Slung Load
In anticipation of the results presented in this paper, it Damping based on inertial stabilized vision control) sys-
is stated that the transportation and accurate deployment tem, which uses an artificial horizon instrument to guide
of loads with single/multiple autonomous aerial vehicles the pilot. The authors are not aware of a completely au-
have been successfully demonstrated in outdoor field ex- tonomous system for full-size helicopters. The system pre-
periments. As was mentioned above, this is a novel feature sented in this work provides the full autonomous control of
that opens the possibility to use aerial robots to assist vic- single-lift, slung-load configurations (based on small-size
tims during rescue phase operations. For instance, it could helicopters), including the active damping of load oscilla-
be possible to command such a transportation system to tions.
deploy medical kits, oxygen masks, satellite phones, etc., The joint transportation of a single load by means of
in places very close to the victims. Accuracy in the deploy- several helicopters has also been proposed in the literature.
ment location is a critical issue in SAR scenarios when in- In experiments with two manned, manually piloted heli-
jured people may have very limited mobility. The system copters, it was determined that the control of two coupled
can be very useful in flooding disaster scenarios in which helicopters is a very challenging task that emphasizes the
hundreds of people are isolated on the roofs of their houses. need for automatic helicopter stabilization. The motivation
The joint transportation of a load by several ground for using two or more small helicopters instead of one with
robots has been a subject of active research and develop- bigger load capacity is as follows:
ment for many years. The coordinated control of the motion
of the vehicles needs to consider the forces involved. Thus, • In the case of real manned transport helicopters, the
each robot could be controlled around a common compli- costs of two small helicopters are often less than for one
ance center attached to the transported object. Assuming with double-load capacity.
that each robot holds the object firmly, the trajectories of • Independent of the load capacity of the most advanced
all the robots determine the trajectory of the object. Both helicopters, there is always a task that requires more
centralized and decentralized compliant motion control al- load capacity than provided by a single helicopter. In
gorithms have been proposed, including the consideration this case the control software allows the coupling of the
of nonholonomic constraints (Kosuge & Sato, 1999). The existing helicopters, in order to form a system with suf-
method has been implemented in an experimental system ficient load capacity.
composed of three tracked mobile robots equipped with a
force sensor. In Sugar and Kumar (1998) the decentralized In particular, research on lifting and transportation of
control of cooperating mobile manipulators is studied, in loads by means of two helicopters (twin-lift, dual-lift) was
which a designated lead robot is responsible for task plan- presented in, e.g., Mittal, Prasad, and Schrage (1991) and
ning. The control of each robot is decomposed (mechan- Reynolds and Rodriguez (1992). This research has been
ically decoupled) into the control of the gross trajectory done only in simulation. Experimental results with a team
and the control of the grasp. The excessive forces due to of aerial robots to manipulate and transport a payload
robot positioning errors and odometry errors are accommo- in three dimensions via cables were recently presented
dated by the compliant arms. In Borenstein (2000) the Omn- (Michael, Fink, & Kumar, 2010). The authors propose two
imate system, which uses a compliant linkage platform be- quality measures for motion plan design that minimize
tween two differential-drive mobile robots, is presented. In individual robot motion and maximize payload stabil-
Huntsberger et al. (2004) the distributed coordinated con- ity along the trajectory. However, field experiments in-
trol of two rovers, carrying a 2.5-m-long mock-up of a pho- volving the lifting and transportation of loads by means
tovoltaic tent is presented and demonstrated as an example of autonomous aerial robots are not addressed in the
of the CAMPOUT behavior-based control architecture. literature.
The single-lift configuration, in which a long rope cou- In this paper, field experiments featuring the load
ples one helicopter and one load, is the only configura- transportation with one helicopter (single-lift) and three
tion commercially utilized for the transportation of slung helicopters (multilift) are presented. Ciolani and Kanning
loads. Several textbooks (see, for example, Wagtendonk, (1992) identify in their work 12 typical slung-load con-
2006) provide information about the correct attachment of figurations: three single-lift configurations, four single-lift
the slung loads and important safety procedures. However, configurations using a twin-rotor helicopter, three dual-
the manual maneuvering of a helicopter with an attached lift configurations, and two multilift configurations. The
slung load is very difficult and requires a skillful and expe- orientation controller proposed in our paper features a
rienced pilot. In particular the active damping of load oscil- measurement-based torque compensation and is applicable
lations is a difficult task that most pilots avoid. Instead the for single-, dual-, and multilift load transportation. There-
pilots stabilize only the helicopter and wait for the load os- fore, the theoretical results of this work are applicable to all
cillation to die down. To support manual piloted slung-load configurations presented by Ciolani and Kanning, except

Journal of Field Robotics DOI 10.1002/rob

916 • Journal of Field Robotics—2011

for the twin-rotor-helicopter configurations, which require link. Owing to the strong magnetic field of the engine, a
an adaptation of the helicopter model used. magnetic field sensor is mounted on the tail.
The described automatic control system allows the use The various payloads necessary for the different exper-
of one or multiple small-size helicopters for slung-load iments were mounted to the same frame, composed of strut
transportation. The number of helicopters is configurable profiles, which was used to attach the components neces-
depending their capabilities and the properties of the load sary for autonomous flight. The following payloads were
to be transported. used during the different missions:
The paper is structured as follows. First, the physical
system is described in Section 2. Then Section 3 is devoted • Fixed visual and infrared cameras.
to the transportation of a single load by means of multiple • A deployment device (DD) developed by the TUB (see
coupled helicopters. Section 4 presents field experiments Figure 2). The functionality of the device is similar to
with one and three helicopters. Finally, the lessons learned the functioning of candy bar vending machines: A metal
from the field experiments and the conclusions close the grommet is attached to a small object (the payload)
paper. through a short wire. This grommet is attached to the
right end of a steel spring. A clockwise rotation of the
spring moves the grommet (and the object) further onto
the spring. This procedure allows the attachment of sev-
2. UNMANNED AERIAL VEHICLE DESCRIPTION
eral small objects to the helicopter. During the drop-
In Figure 1 one of the TUB-H unmanned aerial vehicles ping maneuver the spring rotates counterclockwise until
(UAVs), developed by the Technische Universität Berlin the rightmost grommet is moved beyond the end of the
(TUB) and used for the slung-load transportation exper- spring and the object is released.
iments is shown. The UAVs are based on commercially • The load transportation devices (LTDs), which are specially
available small-size helicopters. The helicopters have a ro- designed for the transportation of slung loads using one
tor diameter of 1.8 m and a main rotor speed of approxi- or more UAVs. The LTD (see Figure 3) is composed of
mate 1,300 rpm and are powered by a 1.8-kW, two-stroke a two-axis cardan joint with two magnetic encoders at-
engine. The UAVs can carry about 1.5 kg of additional pay- tached to each axis. After the joint, a force sensor and
load, and the weight of the UAV itself is 12.5 kg. The dif- the release mechanism for the rope are attached. The
ferent components necessary to achieve autonomous flight release mechanism is composed of a bolt, which is in-
capabilities are mounted to the helicopters, using a frame serted into a tube. The bolt is fixed in the tube through
composed of strut profiles. Through the use of these pro- a pin, which can be pulled out by a small motor to re-
files, the location of hardware components can be altered lease the load. The release mechanism can be used for
and new hardware can be installed easily. This allows quick emergency decoupling of the load from the UAV (and
reconfiguration of the UAVs for different applications, easy for the decoupling from the remaining UAVs), but also
replacement of defective hardware, and alteration of the to release the load after successful transportation. The
positions of different components to adjust the UAV’s cen- magnetic encoders allow the measurement of the rope
ter of gravity. The components necessary for autonomous orientation relative to the UAV fuselage. With this in-
operation are shown in Figure 1: a GPS, an inertial measure- formation and the measured rope force, it is possible to
ment unit (IMU), a control computer, and a communication calculate the torques and forces imposed on the UAV

Magnetic GPS
Compass Antenna

Real-time
GPS computer

Remote
IMU Control
Receiver
WiFi LTD Antenna

Figure 1. TUB-H UAV developed by the TUB used in the field experiments.

Journal of Field Robotics DOI 10.1002/rob

 Bernard et al.: Autonomous Transportation and Deployment with UAVs for SAR Missions • 917

Rotation
Sensor

Node Deployment
Computer Interface
Buttons for
Manual Control

(a) Components of the DD (b) DD loaded with several small objects (sensors in
this photograph)
Figure 2. Detailed view of the DD developed by the TUB. It was used onboard the helicopter for missions that required the
autonomous deployment of objects (sensors) in a given place.

Magnetic Cardan Magnetic

Encoder Joint Encoder

Force
Sensor
Motor

Rope Release
Mounting Pin
Bolt

(a) Components of the LTD (b) TUB-H helicopter equipped with the LTD on the landing pad
Figure 3. Detail of the LTD developed by the TUB. It was used onboard the helicopter for missions that required the autonomous
transportation of objects to a given location.

through the load (and/or the other coupled UAVs). This cess, and the sensor data are presented to the remaining
information is used in the feedback loop of compensator system through the ICSML as generalized objects. The con-
block C; see Section 3.3. troller of the system was designed and tested in simulation
using Matlab/Simulink. The real-time base system (RTBS)
In Figure 4 the concept of the software system run- provides a generic wrapper module to embed C-code gen-
ning on the UAV board computers is shown. The real- erated by Matlab/Simulink. This allows the fast implemen-
time base system is composed of multiple separate mod- tation of new control algorithms and an error-free transition
ules, which are connected by an Interprocess Communi- from Simulink models to real-time controller code. On the
cation and System Management Library (ICSML). The IC- other hand, the executive layer module provides the basic
SML offers, among other features, convenient interprocess functionality needed to integrate the UAV with the rest of
communication methods, using a shared memory design the components of a SAR platform. This includes the com-
and an event system for interprocess notifications (e.g., manding of the UAVs and reporting of the command execu-
on the change of shared memory variables). The driver tion progress and the UAV state. Additionally the executive
modules are used to communicate with the peripheral de- layer module provides a certain level of hardware abstrac-
vices of the UAV, such as the compass, IMU, or GPS. The tion: The relative configuration of the UAVs during the load
device-specific protocol is encapsulated in the driver pro- transportation is part of the control loop and should not

Journal of Field Robotics DOI 10.1002/rob

918 • Journal of Field Robotics—2011

Interprocess
Embedded Communication and

ICSML

ICSML
System Management Executive

MW
Matlab
Code Library (ICSML) Layer

shared
memory

Internal

ICSML
Driver ICSML
ICSML
Driver events Communication
Process
ICSML

Driver
Process Process
Process

Figure 4. RTBS composed of multiple separate modules, which are connected by an ICSML.

be altered by the user of the platform (the coupled UAVs • a higher main rotor revolution speed (than for most full-
are presented as a single entity). This way the command- size helicopters)
ing interfaces for coupled and uncoupled UAVs are equal • a very stiff main rotor without flapping hinges (almost
for both the user and the possible autonomous decision- all purchasable model helicopters)
making software of the SAR platform.
Owing to these differences, the inertial effects of the main
rotor make a significant contribution to the rotational dy-
3. MULTI-UAV LOAD TRANSPORTATION
namics of the system and cannot be neglected. Therefore,
AND DEPLOYMENT
the mechanical model should be considered as being com-
The transportation of loads using only one UAV is strongly posed of two rigid bodies: the main rotor and the fuselage.
limited by the payload capacity of the UAV itself. Assum- For the considered class of model helicopters, the dominant
ing the use of small-size UAVs, this constraint may prevent component is the main rotor and not the fuselage. Unfortu-
the transportation and deployment of loads required for a nately, the main rotor is very often neglected in papers re-
particular application (for example, heavy communication lated to modeling and control of small-size helicopters and
equipment or first aid supplies required for victims in SAR only one rigid body—the fuselage—is accounted for in the
operations). The system designed allows the transportation dynamic equations.
of a single load by means of several helicopters. The num- The complete model of a helicopter is composed of two
ber of helicopters is configurable, depending on the capa- main components: the mechanical model and the model for
bilities of the helicopters and the load to be transported. generation of aerodynamic forces and torques. From exper-
imental results with helicopters we concluded that the gen-
3.1. Modeling eration of aerodynamic forces and torques, at least for the
considered class of helicopters, can be approximated with
The model of a small-size helicopter is a key component for simple algebraic relations (the corresponding time delay
the behavior description of a system composed of one or is approximated by a simple first-order delay). Therefore,
several helicopters, which are coupled to a load by means the dynamics of one small-size helicopter or of a system
of ropes. We use the model presented in our previous composed of several coupled helicopters are mostly deter-
work (Kondak, Bernard, Losse, & Hommel, 2006; Kondak, mined by its mechanical model. The lifting forces FMR and
3
Bernard, Meyer, & Hommel, 2007). A small-size model he- torques TMR generated by the main rotor of each helicopter
licopter shows some specific effects that are not presented 1,2

or are negligibly small in the case of the full-size helicopter (see Figure 5) and the forces FTR2 (not shown in Figure 5)
and vice versa. For this reason, it is impossible to use the generated by the tail rotors are considered to be abstract
models derived for full-size helicopters, see e.g., Johnson control inputs c.
(1980), without any adaptation. As was pointed out in our Figure 5 illustrates the main problem caused by the
previous work, the main differences between model and external coupling of a load to a helicopter (using a rope).
full-size helicopters in respect to modeling and control are The mechanical model of the helicopter is composed of two
as follows: rigid bodies: the fuselage F and the spinning main rotor
MR. The load, denoted as mass point cmL , is connected
• a higher main-rotor-to-fuselage-mass ratio to the helicopter fuselage by means of a rope in the point

Journal of Field Robotics DOI 10.1002/rob

 Bernard et al.: Autonomous Transportation and Deployment with UAVs for SAR Missions • 919

F3MR if helicopter and load are in free fall, Fr and Tr are zero).
MR This creates a more complicated relationship between the
T3MR rotation and the translation dynamics; for each helicopter:
T MR
rotation ⇔ translation.
2
The translational and rotational motion of one partic-
T1MR
ular helicopter has a direct influence on the dynamics of
all other helicopters in the compound, if several helicopters
cm are connected to a common load. Even a translation with
constant acceleration, e.g., to the right in Figure 5, can cause
oscillation of the angle θ between the rope and the heli-
f3 copter axis.
cm

f2 The use of an orientation controller that was designed

r_
p

f1 for an uncoupled helicopter is problematic because of the

r strong mutual coupling between rotation and translation
F (even if the techniques for robust control design are used).
It is noteworthy that in many practical cases the absolute
Fr value of the torque Tr is similar to or even larger than the
values of torque needed to control the rotation of an uncou-
pled helicopter.
n3 For the modeling we consider the general case, in
N n1
which n helicopters are connected to one common load. To
n 2 no derive the dynamic equations of motion we used the Kane
cmL method, see e.g., Kane and Levinson (1985), which allows
us to generate the equations for systems with an arbitrary
Figure 5. Mechanical model of the helicopter connected to the
number of helicopters. The coupling between helicopters
load. It is composed of two rigid bodies: the fuselage F and
and load is established by introducing one motion con-
the spinning main rotor MR. For a real system it is difficult to
straint (the rope length is constant over time) for each heli-
place the point r at the helicopter CoM (cm), which causes the
connecting vector pr -cm to be nonzero.
copter in the compound. The resulting equations are used
for control design as well as for simulation. For more de-
tails on the modeling see Kondak et al. (2007) and Bernard
and Kondak (2010). Currently the aerodynamics of the load
and the ropes are neglected, but in future work, especially
r. The motion of the whole system is considered with re-
for flights with high velocity, they will be considered.
spect to a Newtonian frame N . The point cm is the center of
mass (CoM) of the complete helicopter. For a real system it
3.2. Rope Modeling
is difficult to place the point r at the helicopter CoM cm,
which causes the connecting vector pr -cm to be nonzero. Two assumptions have been made implicitly regarding the
Therefore, the rope force Fr causes a nonzero torque coupling between helicopter(s) and payload during the
Tr = Fr × pr -cm on the helicopter fuselage F . We have modeling: The ropes are rigid and massless and the load
shown in Kondak et al. (2006, 2007) that the rotation dy- transportation device (LTD), required for the estimation of
namics of the helicopter modeled as two rigid bodies is the load position, has no influence on the measurement it-
complicated but not coupled to the translation dynamics. self. However, several real-flight experiments have demon-
Therefore, the equations of the rotational dynamics depend strated that these assumptions are not always valid. As-
only on generalized speeds describing the rotation of the suming normal flight conditions, the ropes are tautened be-
helicopter. The lifting force FMR
3 generated by the main ro- tween helicopter(s) and the payload, similar to a string of a
tor is always approximately perpendicular to the main ro- musical instrument. Therefore, external disturbances (e.g.,
tor plane. This causes the translational accelerations to be wind gusts) or high accelerations can stimulate a natural
completely described by the absolute value of FMR 3 and the oscillation of the ropes. We have observed this behavior in
orientation of the helicopter. Therefore, the relationship be- real-flight experiments several times. The load motion cal-
tween rotation and translation dynamics, for a single heli- culated from angle encoders during these oscillations does
copter, can be expressed as follows: rotation ⇒ translation. not correspond to the real movement of the load. The oscil-
However, if one or several helicopters transport a load lating angles lead to strong responses from the controller
(each helicopter is connected to the load through one rope), and in the worst case to further excitation of the oscilla-
the rope force Fr and the torque Tr act on the fuselage of tions. To analyze these effects, the behavior of LTD, rope,
the helicopter. The torque depends on the orientation of the and payload was studied in theory and in practical experi-
helicopter and its translational motion in the frame N (e.g., ments. First, a model composed of LTD, rope, and payload

Journal of Field Robotics DOI 10.1002/rob

920 • Journal of Field Robotics—2011

1.2 0% 1.2 0%
25% 25%
1 50% 1 50%
Power (normalized)

Power (normalized)
75% 75%
0.8 100% 0.8 100%

0.6 0.6

0.4 0.4

0.2 0.2

0 0
0 5 10 15 20 25 30 35 0 5 10 15 20 25 30 35
Frequency [Hz] Frequency [Hz]
Figure 6. Comparison of simulated (right) and measured (left, load mass 3.0 kg) rope oscillation. The data have been fast Fourier
transformed in order to identify the different oscillation frequencies. The color of the lines denotes the position of the points where
the data were recorded/simulated: blue line, LTD (0%); green/red/cyan lines, 25%, 50%, and 75% of the rope length; purple line,
at the load (100%).

was developed and validated in laboratory experiments. the string l:

Then the model was used to predict the system behavior 
for different system parameters. The model consists of a frope (n) = n( Ft /μ)/(2 l),
series of 20 interconnected pendulums, in which the links where for n = 1 the fundamental eigenfrequency and for
connecting the mass points are considered to be rigid and n ≥ 2 the n th harmonic of the fundamental eigenfrequency
massless. The topmost pendulum models the LTD, and the are calculated. Assuming a load of 3.0 kg, a fundamental
parameters (link length, mass) were chosen to reflect the eigenfrequency of 27.4 Hz is estimated. Third, the peak in
physical properties of the LTD. Similarly, the payload is de- the middle of the figure is a result of the LTD oscillation.
scribed by the bottommost mass point and the mass points The coupling of the LTD to the rope shifts the frequency up-
between LTD and load constitute the flexible rope. ward toward the fundamental eigenfrequency of the rope.
The theoretical model was verified in indoor experi- Based on model predictions, coinciding with our expe-
ments using a pendulum composed of the LTD, a 0.9-m riences from real-flight experiments, the following behavior
rope, and loads of 1.5, 3.0, and 4.5 kg. The pendulum was of the system composed of LTD, flexible rope, and load is
stimulated several times, and the resulting oscillation was estimated: With increasing load mass, the rope fundamen-
recorded by the LTD sensors and a vision system. The vi- tal eigenfrequency becomes much higher than the band-
sion system measured the motion of the LTD and the load, width of the closed-loop system and the helicopter does not
as well as the rope motion at 25%, 50%, and 75% of the respond to the oscillations. However, the rope length has
rope length. Figure 6 shows recorded experimental (left, more influence on the rope eigenfrequency than the load
load mass 3.0 kg) and simulation data (right). As can be mass and therefore for rope lengths of 5/10 m, even for
seen, experiments and simulation match closely and three the high load mass of 4.5 kg, the eigenfrequency becomes
frequency peaks can be identified in the figure: 6.04/3.02 Hz. The helicopter controller will react to these
First, the pendulum frequency can be calculated (ap- oscillations, which can yield instabilities in the system (as
proximated) for the stationary mounting point (high heli- was observed in experiments). As a solution we propose
copter/payload ratios) to be a model-based load motion observer, which will be pre-
      sented in Section 3.4.
fpend = 1/ 2π l/g = 1/ 2π 0.9/9.81 ∼= 0.52 Hz,
3.3. Controller Design
where l describes the rope length and g the gravitational The general scheme of the proposed control algorithm for
acceleration. The actual frequency peak is not visible in the one or several helicopters coupled to a load is composed
figure, because it was truncated during normalization of of two loops: the outer loop for translation control and the
the power spectrum. Second, the rope fundamental eigen- inner loop to control the orientation of each helicopter (see
frequency is approximated using the theory of sound. The Figure 7).
eigenfrequencies of oscillating strings (e.g., a harp string) The input of the control scheme in Figure 7 is the de-
are calculated only from tension Ft , mass μ, and length of sired trajectory x∗ (t) of the helicopters or of the load. The

Journal of Field Robotics DOI 10.1002/rob

 Bernard et al.: Autonomous Transportation and Deployment with UAVs for SAR Missions • 921

The utilization of a force sensor in the rope is a key ele-

Fi 3MR ment of the proposed controller design. The measured rope
system: force is easily recalculated into resultant torques acting on
x* (t ) Fi qi*{1,2} MR
Ti{1,2} the helicopter fuselage. These torques are used in the feed-
1
-
R trans F R ori helis+
back loop of the orientation controller. The use of the force
load sensor signal in the feedback loop has three main advan-
tages:

x(t ) 1. The closed-loop system becomes very robust against

variation of system parameters and disturbances.
Figure 7. General control scheme for one or several heli-
2. The orientation controller design is simplified.
copters coupled to a load. The outer loop performs the transla-
3. The orientation controller does not depend on the num-
tion control, and the inner loop controls the orientation of each
ber of helicopters connected to the load.
helicopter.
As shown in Kondak et al. (2007), the rotation dynam-
ics of a uncoupled helicopter modeled by rigid bodies for
fuselage and main rotor are approximated by the following
translational motion of each helicopter is controlled in the equations:
outer loop by controller Rtrans . Using the deviations from
the desired trajectories the controller Rtrans calculates for T1MR + K12 u2 + K11 u̇1 = 0, (1)
each helicopter i the forces Fi that should be generated by T2MR + K21 u1 + K22 u̇2 = 0, (2)
its rotors. The helicopter i can realize the force Fi by adjust-
ing the absolute value of the main rotor lifting force Fi3 MR MR are the torques generated around the longitudi-
where T1,2
and adjusting the orientation of the main rotor plane or nal and lateral axes of the fuselage, u1,2 are rotation speeds
∗
fuselage (described by the angles qi{1,2} ). The desired ori- of the fuselage, and the coefficients Kxx are constant pa-
entation of the main rotor plane is controlled in the inner rameters of the helicopter and of the main rotor speed. We
loop by controller Rori . The values Fi3 MR and q ∗
i{1,2} are cal- assume that the influence of the rotation around the ver-
culated using algebraic relations in block F−1 . tical axis (rotation speed u3 ) on u1,2 is small and can be
As mentioned above, it is usually impossible to attach considered as disturbance. This assumption is valid if u3
the rope directly to the CoM of the helicopter. This causes is kept close to zero or on constant values. This is achieved
the rope force to create torques, which are imposed on the by an additional tail rotor controller, with a much smaller
helicopter fuselage. Therefore, the influence of the coupled time constant than the time constant of the orientation con-
system disturbs the helicopter orientation and makes the troller for u1,2 . Equations (1) and (2) are coupled through
orientation control a challenging task (Kondak et al. 2006, u1,2 . This coupling leads to high-frequency oscillations (for
2007). The developed approach for the control of coupled the parameters of a typical small-size helicopter) once the
helicopters is based on the following three ideas: system has been stimulated. The control scheme of roll and
pitch angles q1,2 for an uncoupled helicopter is shown in
• The design of the orientation controller (for each he- Figure 8. The blocks shown against a gray background de-
licopter) accounts for the complete dynamics (transla- note the model of the helicopter. The rotation kinematics
tional and rotational) of the whole system, all coupled are represented by block Q, and the rotation dynamics, de-
helicopters and the load. This is required due to the scribed by Eqs. (1) and (2), are represented by block W. The
strong mutual coupling between translation and rota- controller is composed of blocks Q−1 , D and two feedback
tion of each helicopter and the load. loops with gains Ku , Kq for rotation speeds u1,2 and ori-
• The use of the force sensor in the ropes simplifies the entation angles q1,2 , respectively. Desired rotational rates
design of the orientation controller and makes it ro- are calculated from the angular errors e1,2 (interpreted as
bust against variations of system parameters and distur- required Euler angle rates), using controller block Q−1 of
bances. the inverse rotation kinematics. Block D of the controller is
• The design of the translation controller is based on the used to decouple the plant between T1,2 MR and u . The de-
1,2
simplified model and accounts only for the translational coupling can be performed by means of known techniques
dynamics of the whole system. In this simplified model from linear control theory, e.g., using a matrix composed
the helicopters and the load are modeled as mass points. of compensating transfer functions. This orientation con-
The model allows the application of arbitrary forces on troller shows good performance and robustness in simu-
each mass point, which represents a helicopter. How- lation and real-flight experiments for different types of he-
ever, the generation of these forces is subjected to a lin- licopters, as described in Kondak et al. (2006, 2007).
ear dynamic model, which reflects the force generation As discussed above, the motion of the whole system
process of the helicopters. has a strong influence on the rotational dynamics of the

Journal of Field Robotics DOI 10.1002/rob

922 • Journal of Field Robotics—2011

Tr 2
Fr
q1,2,3 C Tr1

Ku

*
q1* - e1 - u1 T1 - u1 u1 q1 q1
Kq
1
Q D W Q
q2* e2 u2* T2 - u2 u2 q2 q2
Kq
- -
u6
Ku

Figure 8. Scheme for the orientation control. The resulting controller is composed of the orientation controller for an uncoupled
helicopter and compensator block C.

coupled helicopters. To account for this influence, block D lage is measured through Fr . Second, as long as the orien-
should be replaced by the inverse rotational dynamics D̃ tation of the helicopter is known, the calculated compensa-
not of a single helicopter, but of the whole system (consid- tion torque is always in the correct phase.
ering both the rotation and translation of each helicopter). The details of the presented control algorithms can be
Utilizing this new block D̃, the orientation controller for a found in Bernard and Kondak (2009, 2010) and Kondak
helicopter coupled to a load shows performance equal to et al. (2006, 2007).
that of the orientation controller with block D for an un-
coupled helicopter, considering nominal values of all sys-
tem parameters ρi . 3.4. Load Motion Observer
It was shown in simulation that the orientation con- Oscillations of the LTD and rope disturb the motion es-
troller based on inversion block D̃ for a coupled helicopter timation of the load, as explained in Section 3.2. There-
(unlike block D for a uncoupled helicopter) is quite sen- fore, a direct estimation of the load motion based on the
sitive to variation of the system parameters ρi (5% varia- measured LTD angles is not feasible. The expected oscilla-
tion could be critical). To overcome this problem, we pro- tion frequency of the undisturbed load motion is very low
pose to use a force sensor in the rope. The measured rope (0.52 Hz assuming a rope length of 0.9 m). Therefore, a very
force Fr is used to calculate the influence of the coupled simple solution, the low-pass filtering of the measured an-
system on the rotational dynamics of the helicopters. This gles, was applied. Figure 9 shows an angle measured by
influence is expressed by the torque Tr = Fr × pr -cm , where the LTD (black line) and the angle after application of a
pr -cm is the position vector, connecting the rope attachment fourth-order Butterworth low-pass filter (gray dashed line,
point r and the helicopter CoM. The resulting orientation cutoff frequency fc = 2 Hz). The gray line corresponds to
controller is composed of the orientation controller for an the result of an observer and is explained later. The dis-
uncoupled helicopter and a compensator block C (see Fig- turbance (between 3 and 5 s) still has strong influence on
ure 8). The block estimates the disturbance torques Tr1 , Tr2 the filtered angle. Additionally, the filtered angle clearly ex-
and subtracts them from the torques calculated by the ori- hibits a large time delay if it is compared to the measure-
entation controller. The use of the compensator C allows us ment.
to decouple the orientation control for each helicopter from Therefore, an observer-based estimation has been im-
the rest of the system. This makes the use of the same ori- plemented and tested. First we used a linear observer based
entation controller, independent of the number of coupled on the flexible rope model, described in Section 3.2, with a
helicopters, possible. high number of mass points for the rope approximation.
There are two reasons for the robustness of the pro- This observer converges quickly to the state of the simu-
posed orientation controller: First, the compensation is in- lated model (the model is observable), but unfortunately
dependent of the mass of the load and the length of the it is very sensitive to parameter variations. Assuming that
rope, because the actual influence of the load on the fuse- the rope length differs by 0.1 m or the load mass differs by

Journal of Field Robotics DOI 10.1002/rob

 Bernard et al.: Autonomous Transportation and Deployment with UAVs for SAR Missions • 923

Figure 9. Comparison of simulated and measured rope oscillation. The black line represents an angle measured by the LTD, the
gray dashed line is the angle after application of a fourth-order Butterworth low-pass filter (cutoff frequency fc = 2 Hz), and the
gray line corresponds to the result of the developed observer.

0.1 kg, the observer is not converging. Therefore, a different European Commission (June 2006–September 2009). The
approach was tested: The observer was calculated using the researchers from the University of Seville led a consortium
simplified model for the controller design. In this model the of five Universities (TUB and the Universities of Seville,
ropes are simplified to be rigid and massless. The result- Bonn, Stuttgart, and Twente) and three companies (Selex
ing observer is robust against parameter variation and con- Sensors and Airborne Systems, Flying-Cam, and the Iturri
verges quickly to the motion of the undisturbed pendulum. Group) from five European countries.
Movements besides the pendulum frequency of 0.52 Hz are The general objective of the project was the de-
effectively filtered out by the observer. Once again Figure 9 sign, development, and demonstration of a platform com-
is used to compare the performance of the observer to the posed of heterogeneous systems that are able to operate
filter and the direct measurement. The angle estimated by in a distributed manner in disaster management scenar-
the observer exhibits less sensitivity to the disturbance (be- ios without preexisting (or with damaged) infrastructure
tween 3 and 5 s) than the filtered angle. Additionally, the (Maza, Caballero, Capitan, de Dios, & Ollero, 2011). Thus,
observed angle has only a small time delay. In fact, after the the platform comprises self-deployment capabilities (in
observer converged, no time delay was visible compared to particular, the autonomous transportation and deploy-
the measured angle. ment of different load types, such as small sensors, cam-
The observer was tested several times in real-flight eras, and communication equipment) using one or sev-
experiments. Strong, steady winds and wind gusts were eral small-size helicopters. The systems integrated in the
present during the experiments and demonstration con- platform include aerial robots, wireless sensor networks,
ducted in May 2009 (see Section 4). Despite these bad ground-fixed-cameras, and ground vehicles with actuation
environmental conditions several load transportation ex- capabilities.
periments were conducted successfully. To cross-check its The TUB-H UAVs were used several times in differ-
functioning, one experiment was conducted without an ob- ent experiments, with various payloads such as infrared
server, using the angles directly measured by the LTD. The and/or visual cameras, the DD, and the LTD. Configura-
experiment needed to be aborted almost immediately be- tions with single and multiple UAVs were tested for the
cause the rope oscillations caused a very strong reaction of transportation and deployment of objects in SAR scenar-
the controller. These experiments proved the necessity and ios. A selection of these field experiments is described in
the functional efficiency of the observer. the following.

4. VALIDATION IN FIELD EXPERIMENTS 4.1. Sensor Dropping

The work presented in this section was conducted within During SAR operations the deployment could be required
the framework of the AWARE Project1 funded by the of additional light objects, such as small sensors or com-
munication devices in specific locations around the affected
area. Within the framework of the AWARE project, a num-
1
http://www.aware-project.net. ber of missions involving the deployment of small wireless

Journal of Field Robotics DOI 10.1002/rob

924 • Journal of Field Robotics—2011

(a) One TUB-H autonomous helicopter (b) The hovering above the deployment position begins at 160 s and ends at 300 s;
dropping a sensor during this time span the UAV descends, deploys the node, and ascends again, while
maintaining the hovering position in the x/y plane, with a precision of 0.4 m
Figure 10. Sensor-dropping operation by the TUB-H autonomous helicopter equipped with the DD.

sensors were carried out. In Figure 10, the sensor deploy- measured, which introduced repeated displacement of the
ment using one TUB-H UAV is shown. The hovering above load.
the deployment position begins at 160 s and ends at 300 s. During the experiment the load position observer was
During this time span the UAV descends, deploys the node, successfully tested. The position of the load is required in
and ascends again, while maintaining the hovering posi- order to stabilize the load and is calculated using the angles
tion in the x/y plane, with a precision of 0.4 m. In the z axis measured by LTD. During flight, the rope is normally taut
a deviation of 1 m was recorded, which is more than twice between the helicopter fuselage and the load. If animated
the deviation of the x and y axes. The reasons were wind by external influences (e.g., by wind gusts) the rope be-
gusts of 50 km/h occurring during the flight. In these ex- gins to oscillate like the string of a musical instrument. The
periments the gain of the altitude controller was reduced in purpose of the load observer is to estimate the position
order to reduce mechanical stress on the helicopters flying of the load while rejecting oscillations introduced into the
under such harsh conditions. As a result, the winds gusts rope itself. Two experiments were conducted: the first with-
are able to lift the helicopter up or push it down about 1 m, out the observer, to prove that these rope oscillations occur
until the height controller is able to compensate. during real-flight experiments, and the second with the ob-
server, to prove that the load position observer provides a
working solution to the problem. Although the first exper-
iment needed to be aborted, due to the strong controller
4.2. Slung-Load Transportation Using a Single UAV
response to the oscillation of the rope, in the second exper-
The flight data shown in Figure 11 were recorded on 27 May iment the observer damped the rope oscillations and pre-
2009 in Spain. The plot shows a flight step of 5 m and one served the motion of the load.
of 15 m; another step of 30 m was omitted in order to avoid During hovering, wind gusts of 30 km/h led to
detail loss (caused by the coarse scaling of the axis). For damped oscillations of the load, with maximum amplitude
this experiment a single helicopter equipped with the LTD of 0.5 m. For real-world applications this is a good value,
is used. A jerry can, connected to the LTD using a rope, is considering the weather conditions. Owing to the bowing
transported by the helicopter over a distance of 5 and 15 m. of the rope caused by wind, only an approximate estima-
The rope was 5 m long, and the weight of jerry can was tion of the absolute load position is possible. For example,
1.1 kg. During the experiment, wind gusts of 30 km/h were

Journal of Field Robotics DOI 10.1002/rob

 Bernard et al.: Autonomous Transportation and Deployment with UAVs for SAR Missions • 925

16
UAVx
UAVy
14
UAV z
load x
12 load y
load z

10

position [m]
8

-2
0 20 40 60 80 100 120 140 160 180 200 220
time [s]
(a) Single TUB-H autonomous heli- (b) The plot shows a flight step of 5 m and one of 15 m; another step of 30 m was
copter transporting a load omitted in order to avoid detail loss (caused by the coarse scaling of the axis)
Figure 11. Slung-load transportation using one UAV.

an error of 5 deg, caused by the bowing of the rope, causes a licopters during the whole flight are shown in Figure 12.
deviation 0.44 m from the real position (for a rope length of The xh123 and yh123 coordinates describe the horizontal mo-
5 m). For real-world applications two solutions exist: first, tion of the helicopters, and the zh123 coordinates describe
the use of additional sensors to estimate the position of the the vertical motion. The coordinates of the helicopters are
load (e.g., vision-based estimation) and second, the manual
placement of the load by the UAV operator, in which the op- 20
erator changes the UAV position step by step until the load
zh123
is directly above the desired position. The manual place- 15
ment is favorable whenever the exact GPS position of the
placement location is unknown and cannot be measured 10
x
h123

before takeoff. In that case the helicopter operator uses vi-

sual feedback of the deployment process to maneuver the
distance [m]

5
load to the desired position.
0
4.3. Load Transportation with Multiple UAVs
−5
The first successful experiment was conducted by the au- y
h123
thors in Berlin, December 2007.2 For these flight experi-
−10
ments three identical helicopters as described in Section 2
were used. The helicopters are equipped with a multi-UAV load on the ground load in the air
−15
modular autopilot system developed at TUB. The rope is 0 20 40 60 80 100
time [s]
120 140 160 180 200

attached to the helicopter by means of the LTD, which is

Figure 12. Motion of three helicopters transporting one load
mounted between the landing skids. A load of 4 kg was
of 4 kg. The load was lifted when the helicopters reached ap-
transported by three helicopters. In this experiment ropes
proximate 12.4 m. The coordinates of the helicopters are shown
with a length of 13 m were used. The helicopters were ar- in different ground-fixed frames that have the same orientation
ranged as an equilateral triangle on the ground, with a dis- but different origins (takeoff position of each helicopter); there-
tance of 8 m between them. The coordinates of all three he- fore there are no offsets between the helicopter trajectories. The
xh123 and yh123 coordinates describe the horizontal motion of
the helicopters, and the zh123 coordinates describe the vertical
2
http://www.youtube.com/watch?v=tl6DYWNe9ac. motion.

Journal of Field Robotics DOI 10.1002/rob

926 • Journal of Field Robotics—2011

shown in different ground-fixed frames that have the same the first field experiment involving the transportation of a
orientation but different origins (takeoff position of each load from the ground to the roof of a building with three
helicopter); therefore there are no offsets between the he- autonomous helicopters. Several videos with the live exe-
licopter trajectories. The mapping of helicopters and line cution of the mission are included in the online version of
colors is ambiguous. This is acceptable because the mo- this article.
tions of the helicopters are discussed together and not The experiment was executed as follows: The load
individually. transportation system (LTS), composed of three TUB-H he-
The load was lifted when the helicopters reached ap- licopters and the coupled payload, was ready for opera-
proximately 12.4 m. The weight of the load was not con- tion, and the helicopters were waiting on the three takeoff
sidered in the controller, and therefore a small disturbance and landing pads. The platform user specified a load trans-
in the zh123 trajectories can be observed at the moment the portation task to deploy the wireless pan–tilt camera on the
load was lifted from the ground as well as during strong top floor, and a plan builder module generated the full set
acceleration in the x, y direction. A position error of each of ordered tasks for the LTS.
helicopter in hovering was approximately ±0.3 m. During It should be noted that the enormous complexity of the
the whole flight the triangular formation of the helicopters LTS composed of three helicopters was hidden to the user,
with a precision of about ±0.3 m and the load was moved who needed only to specify the deployment task by pro-
very smoothly. To our knowledge these were the very first viding the desired GPS location of the load deployment.
successful flight experiments to transport a load using mul- The altitude specified for the deployment was several me-
tiple autonomous helicopters. ters above the top floor of the building. The autonomous
Another experiment was carried out in May 2009 decision-making software of the platform had access to the
within the scope of the AWARE disaster scenario demon- map of the area in order to plan the deployment task de-
stration: A fire alarm had been declared in a building, and composition properly, also taking into account the length
the objective was to place a wireless camera with pan–tilt of the ropes.
on the top floor. The camera could provide continuous real- Figure 13 shows a two-dimensional (2-D) view of the
time video to monitor the operations of the firemen and trajectories followed by the three TUB-H helicopters and
the health status of the victims on the roof of the building. the transported pan–tilt camera unit. The formation of the
Several coupled helicopters were required for this task be- helicopters was similar to the formation used for the ex-
cause the camera, together with its associated communica- periment conducted in December 2007: an equilateral tri-
tion equipment and batteries, was too heavy for a single angle, with a distance of 8 m between the helicopters. The
helicopter. To the best of the authors’ knowledge, this was weight of the payload composed of camera, transportation

Figure 13. Path followed by the three helicopters transporting the load in the x–y plane. The trajectories of the load and the
helicopters are in black and gray, respectively. The arrow indicates the direction of the load transportation. The formation of the
helicopters was an equilateral triangle, with a distance of 8 m between the helicopters.

Journal of Field Robotics DOI 10.1002/rob

 Bernard et al.: Autonomous Transportation and Deployment with UAVs for SAR Missions • 927

Figure 14. Values of the x, y, and z coordinates of the helicopters (left) and the load (right) during the flight. The coordinates of
each helicopter and the load are given relative to one global Newtonian reference frame, but for each helicopter and for the load
the coordinates were plotted with respect to a relative origin (the takeoff position of the respective helicopter or the liftoff position
of the load). The trajectories of the helicopters show almost no deviation relative to each other.

housing, and batteries was approximately 5 kg. The trajec- It should be mentioned that during the execution, wind
tories of the helicopters are plotted in gray, and the trajec- gusts around 35 km/h were registered. Therefore, the per-
tory of the camera is plotted in black. The estimated rope formance of the coupled LTS is considered to be very good.
positions are plotted as gray lines, where a time step of Several photographs of the LTS during the execution
1 s elapsed between two successive estimations. The ar- of the mission are shown in Figure 15. A fourth helicopter
row indicates the direction of the load transportation. It is used to acquire airborne video footage of the mission can be
noteworthy that despite the stormy weather conditions, the seen in two pictures. Finally, Figure 16 contains a screenshot
load stayed clearly within the extends of the roof during the of the human–machine interface application captured dur-
deployment. ing the execution, as well as one image transmitted from
Figure 14 complements the flight data presented in the transported camera once it had been deployed on the
Figure 13, by depicting the values of the x, y, and z coor- top of the building.
dinates of the helicopters and the load during the flight.
The motion of the load is depicted from the moment the
load was lifted off the ground to the moment it was de-
ployed on the roof. The coordinates of each helicopter and 5. CONCLUSIONS AND LESSONS LEARNED
the load are given relative to one global Newtonian refer- The experiments conducted within the AWARE demonstra-
ence frame, but for each helicopter and for the load, the tion in 2009 imposed high availability requirements on all
coordinates were plotted with respect to a relative ori- involved subsystems. For two reasons this was particularly
gin (the takeoff position of the respective helicopter or the true for the autonomous helicopters: First, the helicopters
liftoff position of the load). The mapping of helicopters and were the designated platform for the transportation (and
line colors is ambiguous, but similar to Figure 12 the mo- deployment) of various sensors and therefore they were
tions of the helicopters are only discussed together. The tra- involved in almost every experiment. Second, the exper-
jectories of the helicopters show almost no deviation rela- iments included the demonstration of the LTS, which re-
tive to each other. Therefore, the helicopters preserved their quires the simultaneous operability of three helicopters.
relative formation during the whole flight. To demonstrate Therefore, several preventive measures were taken to en-
the robustness of the translation controller, the additional sure the successful conduct of the experiments. The com-
weight of the load was neglected during the experiment. plexity of these measures ranged from simple (such as the
Therefore, in the interval [120 s, 140 s] the strong accelera- regular maintenance of the helicopters or the availability of
tion of the helicopters caused a height deviation. However, spare parts) to complex (such as the adherence to preflight
besides this deviation no additional side effects were no- checks/procedures or the possibility to adapt hard- and
ticeable. During the return flight [240 s, 260 s], after the software in the field, toward the requirements of the experi-
load deployment, no disturbance of the height is visible. ment or the actual environmental conditions). In particular,

Journal of Field Robotics DOI 10.1002/rob

928 • Journal of Field Robotics—2011

Figure 15. Three UAVs transporting a pan–tilt wireless camera to the roof of a building with a height of 12 m in Utrera (Spain) in
May 2009. The images show the mission during the takeoff (top), during the actual load transportation (middle left), shortly before
(middle right), and shortly after (bottom) the load deployment. In two pictures a fourth helicopter is visible, which was used to
acquire airborne video footage of the mission.

Journal of Field Robotics DOI 10.1002/rob

 Bernard et al.: Autonomous Transportation and Deployment with UAVs for SAR Missions • 929

(a) Map of the area with the position and heading of the three LTS helicopters represented by arrows; images transmitted
by a fourth helicopter and telemetry from all the UAVs

(b) Camera deployed on the top floor of the building; the op-
erator used the pan–tilt to find one victim; the LTS is still
over the deployment location after releasing the ropes
Figure 16. Screenshot of the platform human–machine interface during the execution of the load transportation mission in 2009.
The different elements in the interface were (top left) map of the area with the position and heading of the three LTS helicopters
represented by arrows; (center) images transmitted by another helicopter and telemetry from all the UAVs; (right) interface to
control the transported camera with pan–tilt.

the adaptability of hardware and software has been proven quick relocation of the payload to a different helicopter and
to be useful during the experiments: the resumption of the experiments.
The combustion engine of a helicopter exhibited signs The adverse weather conditions with high wind
of imminent failure during the conduct of an experiment. speeds (≥25 km/h) and strong wind gusts (≥35 km/h),
Therefore, the experiment was aborted immediately and particularly during the last days of the AWARE project
the helicopter was landed safely. Then, in order to proceed experiments, required the adaption of the helicopter con-
with the experiment as quickly as possible, the payload was troller. It was possible to operate the helicopters during
mounted to a different helicopter. The modular design of these bad weather conditions, although the controller was
the system components, in combination with the strut pro- optimized for the operation during medium wind speeds
file mounting frame (described in Section 2), allowed the (≤20 km/h). The controller was able to quickly compensate

Journal of Field Robotics DOI 10.1002/rob

930 • Journal of Field Robotics—2011

for the displacements of the helicopter (caused by the wind

gusts). However, the helicopter was operating close to its Extension Media type Description
performance limit, and the strong reactions of the con-
troller caused unnecessarily high stress on rotors, structure, 1 Video Load transportation mission with
multiple UAVs (ground camera
and engine. Therefore, the control coefficients of the height
view)
controller were relaxed, which allowed the helicopter to
2 Video Load transportation mission with
be lifted up or to be pushed down (±1 m) by the wind
multiple UAVs (transported
gusts. This measure strongly reduced the stress on the he-
wireless camera view)
licopter and improved the precision of the remaining two 3 Video Load transportation mission with
position coordinates. In Section 4.1 an experiment utilizing multiple UAVs
the modified controller coefficients is described. The con- (human–machine interface
troller coefficients were modified online during a single test view)
flight, in which the Matlab/Simulink external mode inter- 4 Video AWARE project summary
face was utilized; see Section 2 for a description of the soft- (including load transportation)
ware system. Without the possibility of online parameter
modification, several test flights would have been required
until good parameters could be found. Using the Mat-
lab/Simulink interface it was possible to upload different ACKNOWLEDGMENTS
controller coefficients and directly monitor the improve- This work was partially supported by the AWARE Project
ments achieved by a particular set of coefficients. Therefore, (IST-2006-33579) funded by the European Commission un-
only a single flight of approximately 10-min length was re- der FP6 and also by the ROBAIR Project (DPI2008-03847)
quired for the adaption of the coefficients, and the schedule funded by the Spanish Research and Development Pro-
of planned experiments was only insignificantly delayed. gram.
The above examples were chosen because they clearly
demonstrate the benefits of the excellent system adaptabil-
ity. However, beside these examples, the importance of the REFERENCES
system adaptability was witnessed several times during the
field experiments. Therefore it is possible to conclude that Arai, M., Tanaka, Y., Hirose, S., Kuwahara, H., & Tsukui, S.
adaptability of the system was a key element required for (2008). Development of “Souryu-Iv” and “Souryu-V”: Se-
the successful completion of the field experiments. rially connected crawler vehicles for in-rubble searching
The paper has presented the multi-UAV LTS, which operations. Journal of Field Robotics, 25(1–2), 31–65.
requires consideration of physical interactions between Bernard, M., & Kondak, K. (2009, May). Generic slung load
aerial robots. The multi-UAV architecture developed in the transportation system using small size helicopters. In Pro-
ceedings of the IEEE International Conference on Robotics
AWARE project allows different levels of interaction among
and Automation, Kobe, Japan (pp. 3258–3264).
the UAVs and between the UAVs and the environment, in-
Bernard, M., & Kondak, K. (2010). Load transportation system
cluding both sensing and actuation. In particular, the pa-
based on autonomous small size helicopters. Aeronautical
per has presented results obtained in the AWARE project Journal, 114(1152), 191–198.
demonstrating the lifting and transportation of a slung load Birk, A., & Carpin, S. (2006). Rescue robotics—A crucial mile-
by one helicopter and also by three coupled helicopters, stone on the road to autonomous systems. Advanced
which has been the first demonstration of this challenging Robotics Journal, 20(5), 595–695.
application. Borenstein, J. (2000). The OmniMate: A guidewire- and beacon-
The proposed methods open many different new op- free AGV for highly reconfigurable applications. Inter-
portunities in missions involving the cooperation of multi- national Journal of Production Research, 38(9), 1993–
ple UAVs for applications such as SAR and interventions in 2010.
disaster management and civil security. The transportation Cicolani, L. S., & Kanning, G. (1992). Equations of motion
of loads by means of UAVs can be also considered as a first of slung-load systems, including multilift systems (Tech.
step toward cargo transportation by UAVs or even toward Rep. NASA TP-3280). NASA Ames Research Center.
the evacuation of people in SAR operations. Corke, P., Hrabar, S., Peterson, R., Rus, D., Saripalli, S., &
Sukhatme, G. (2004a, April/May). Autonomous deploy-
ment and repair of a sensor network using an unmanned
6. APPENDIX: INDEX TO MULTIMEDIA EXTENSIONS aerial vehicle. In Proceedings of the IEEE International
Conference on Robotics and Automation, New Orleans,
The table shows the different videos of the LTM carried out LA (pp. 3602–3608).
in 2009 in the framework of the AWARE Project. The videos Corke, P., Hrabar, S., Peterson, R., Rus, D., Saripalli, S., &
are available as Supporting Information in the online ver- Sukhatme, G. (2004b, June). Deployment and connectivity
sion of the article. repair of a sensor net with a flying robot. In Proceedings

Journal of Field Robotics DOI 10.1002/rob

 Bernard et al.: Autonomous Transportation and Deployment with UAVs for SAR Missions • 931

of the 9th International Symposium on Experimental Conference on Robotics and Automation, Rome, Italy
Robotics, Singapore. (pp. 2375–2380).
Gage, A., Murphy, R., Rasmussen, E., & Minten, M. B. (2004). Kosuge, K., & Sato, M. (1999, October). Transportation of a sin-
Shadowbowl 2003: Lessons learned from a reachback ex- gle object by multiple decentralized-controlled nonholo-
ercise with rescue robots. IEEE Robotics & Automation nomic mobile robots. In Proceedings of the IEEE Interna-
Magazine, 11(3), 62–69. tional Conference on Intelligent Robots and Systems, Ky-
Goodrich, M. A., Morse, B. S., Gerhardt, D., Cooper, J. L., ongju, Korea (vol. 3, pp. 1681–1686).
Quigley, M., Adams, J. A., & Humphrey, C. (2008). Sup- Maza, I., Caballero, F., Capitan, J., de Dios, J. M., & Ollero, A.
porting wilderness search and rescue using a camera- (2011). A distributed architecture for a robotic platform
equipped mini UAV. Journal of Field Robotics, 25(1–2), with aerial sensor transportation and self-deployment ca-
89–110. pabilities. Journal of Field Robotics, 28(3), 303–328.
Hollinger, G., Singh, S., Djugash, J., & Kehagias, A. (2009). Effi- Michael, N., Fink, J., & Kumar, V. (2010). Cooperative manipu-
cient multi-robot search for a moving target. International lation and transportation with aerial robots. Autonomous
Journal of Robotics Research, 28(2), 201–219. Robots, 30(1), 73–86.
Hsieh, M. A., Chaimowicz, L., Cowley, A., Grocholsky, B., Micire, M. J. (2008). Evolution and field performance of a res-
Keller, J. F., Kumar, V., Taylor, C. J., Endo, Y., Arkin, R. C., cue robot. Journal of Field Robotics, 25(1–2), 17–30.
Jung, B., Wolf, D. F., Sukhatme, G., & MacKenzie, D. C. Mittal, M., Prasad, J. V. R., & Schrage, D. P. (1991). Nonlin-
(2007). Adaptive teams of autonomous aerial and ground ear adaptive control of a twin lift helicopter system. IEEE
robots for situational awareness. Journal of Field Robotics, Control Systems Magazine, 11(3), 39–45.
24(11), 991–1014. Murphy, R., Tadokoro, S., Nardi, D., Jacoff, A., Fiorini, P.,
Hsieh, M. A., Cowley, A., Kumar, V., & Taylor, C. J. (2008). Choset, H., & Erkmen, A. (2008). Search and rescue
Maintaining network connectivity and performance in robotics. In Springer Handbook of Robotics (pp. 1151–
robot teams. Journal of Field Robotics, 25(1–2), 111– 1173). Berlin: Springer Verlag.
131. Murphy, R. R., & Stover, S. (2008). Rescue robots for mudslides:
Huntsberger, T. L., Trebi-Ollennu, A., Aghazarian, H., A descriptive study of the 2005 La Conchita mudslide re-
Schenker, P. S., Pirjanian, P., & Nayar, H. D. (2004). Dis- sponse. Journal of Field Robotics, 25(1–2), 3–16.
tributed control of multi-robot systems engaged in tightly Poppinga, J., Birk, A., & Pathak, K. (2008). Hough based terrain
coupled tasks. Autonomous Robots, 17(1), 79–92. classification for realtime detection of drivable ground.
Johnson, W. (1980). Helicopter theory. Mineola, NY: Dover Journal of Field Robotics, 25(1–2), 67–88.
Publications. Reynolds, H. K., & Rodriguez, A. A. (1992). H∞ control of
Kane, T., & Levinson, D. (1985). Dynamics: Theory and appli- a twin lift helicopter system. In Proceedings of the 31st
cations. New York: McGraw-Hill. IEEE Conference on Decision and Control, Tucson, AZ
Kondak, K., Bernard, M., Losse, N., & Hommel, G. (2006, May). (pp. 2442–2447).
Elaborated modeling and control for autonomous small Sugar, T., & Kumar, V. (1998). Decentralized control of coop-
size helicopters. In ISR/ROBOTIK 2006 Joint Conference erating mobile manipulators. In Proceedings of the IEEE
on Robotics, Munich, Germany. International Conference on Robotics and Automation,
Kondak, K., Bernard, M., Meyer, N., & Hommel, G. (2007, Leuven, Belgium (vol. 4, pp. 2916–2921).
April). Autonomously flying VTOL-robots: Modeling Wagtendonk, W. (2006). Principles of helicopter flight. New
and control. In Proceedings of the IEEE International Castle, WA: Aviation Supplies & Academics.

Journal of Field Robotics DOI 10.1002/rob