Multi-robot Systems Localization Using

Cooperative SLAM
A research proposal

I.

INTRODUCTION

In comparison with fixed robots who always have a place

in manufacturing, mobile robots promise to work more
flexibly and they are friendlier to end-users in many
applications. These applications include medical and
surgical uses, personal assistance, security, warehouse and
distribution applications, as well as ocean and space
exploration. Mobile robots are capable of traveling across
their environment, therefore, the possibility of interacting
with other robots in their surroundings arises. It is very
exciting to study how multiple robots could cooperate
towards reaching a common goal. In addition, teams of
robots can handle tasks that are difficult or even impossible
for a single robot or they may accomplish duties in a faster
or cheaper way. For example, robots with different sensing
capabilities can use relative measurements to help another
robot who in their communicating range or other agents in
team to localize more accurately. In reality, many
organizations have developed multi-robot systems to carry
on some repeated tasks or work in dangerous environments
where humans are unable to appear. For instance, Defense
Advanced Research Project Agency (DARPA) in USA have
developed the Autonomous Vehicle Aerial Tracking and
Reconnaissance (AVATAR) for military surveillance,
DARPA Perception of Off road Robotics (Percept OR)
providing an autonomous flying eye helicopter and
Mobile Autonomous Robot Software (MARS) for rescuing
[9].
In order to accomplish these mentioned duties and benefit
human beings, it is important for each agent in multi-robot
systems is able to localize itself and help other team
members on localization. This is known as cooperative
localization (CL) [8]. This localization approach can be
classified into three main categories based on estimation
architecture: (i) centralized CL approaches, (ii) multicentralized CL approaches, and (iii) decentralized CL
approaches [5]. In [8], the authors have stated a fourth
category of CL approach named distributed deader-assistive
localization method. While centralized methods perform
better than decentralized counterparts [4], these approaches
have some limitations, saying high computational and
communicational cost. Thus, decentralized approaches
attract more attention form researchers all over the world
and have got some significant achievements. However, the
problems with this CL methods, namely consistent
distributed estimation and Robust distributed perception
(data association) [12] are not thoroughly solved. The
proposal presented in this paper aims to investigate in
decentralized cooperative localization and mapping
approaches in order to achieve the better performance of

robot localization in unknown environment and tackle the

challenges listed above.
II.

LITERATURE REVIEW AND PROBLEM

STATEMENT

There are many methods for robot localization. Therefore,

before
thoroughly investigating into
cooperative
localization, it is necessary to have a look at what
researchers have done in order to deal with this topic. In
[11], the authors have mentioned about classical beaconbased localization algorithms which was introduced in [13].
The method employed the extended Kalman filter (EKF) to
solve the problem of mobile robot localization in a known
environment. The algorithm of this approach relied on the
concept of geometric beacon a naturally occurring
environment feature that can be reliably observed in
successive sensor measurements and can be accurately
described in terms of a concise geometric parameterization.
The algorithm is based on an EKF that utilizes matches
between observed geometric beacon and a prior map of
beacon locations to maintain a robust vehicle location
estimation. However, this method may be inapplicable in
practical environment because it is dynamic, cluttered and
unpredictable.
Simultaneous Localization and Mapping (SLAM), also
known as Concurrent Mapping and Localization, is one of
the most popular methods for a single robots localization in
indoor environments [3]. SLAM problems arise when the
robot does not have access to a map of the environment; nor
does it have access to its own poses. Instead, all it is given
are measurements z1:t and controls u1:t. In SLAM, the robot
acquires a map of its environment while simultaneously
localizing itself relative to this map. From a probabilistic
perspective, there are two main forms of the SLAM problem
[2], which are in the same level of importance in practice.
The first is called online SLAM problem: It involves
estimating the posterior over the momentary pose along with
the map:
p(xt, m | z1:t, u1:t)
where xt is the pose at time t, m is the map, and z 1:t and u1:t
are the measurements and controls, respectively.
The second SLAM problem is called the full SLAM
problem. In full SLAM, it is seek to calculate a posterior
over the entire path x 1:t along with the map, instead of just
the current pose xt:
p(x1:t, m | z1:t, u1:t)
Although SLAM methods, such as laser-based or visionbased ones, can help the robots be able to cope with
unknown environments when localizing, the high processing
and payload prevents applications of these methods. In

addition, SLAM does not implement missions effectively

when the environments are changing. Thus, this paper are
written to propose the method for better performance of
SLAM, which will be discussed in details later.
GPS (Global Positioning System) is another widely used
methods in robot localization. The methods are highly
effective in outdoor environments, but urban canyon would
be a challenge for this approach. Besides, for the systems
designed to work in dynamic environments, saying indoor
and outdoor, the GPS surely is unable to fulfil the localizing
tasks.
Due to the existing limitations of these approaches for robot
localization many robotics researchers all over the world
have paid more attention on cooperative localization in
recent years. The Intelligent LAB also have a considerable
number of publications on the topic. In [3], the multi-robot
systems is developed to enhance localizing ability of
individual robot while reducing the cost of the entire system.
In order fulfil these problems, the relative localization
scheme is employed. The whole system have only one
leader robot being able to localize itself in global frame. The
other child robots including both ground and air types are
only able to localize themselves in the measurement range
with the leader robot. After fusing the estimation, the child
robots can use the leader robots belief to improve their pose
estimation in the world frame. This method can decrease the
cost of the child robot, being equipped no means of
localizing themselves in global coordinate, but the
effectiveness of localization still have been achieved.
However, this approach still maintains some limitations: (i)
the child robot have to operate in sensing field of the leader
and all the robots have to carry on their missions within the
communicating range of the remote host; (2) the leader
robot cannot use the relative measurements between the two
child robot during localization process.
The limitations of the method in [3] are eliminated in the
method introduced in [8]. The distributed leader-assistive
system contains two groups of robots. Robot with higher
sensor payload, higher processing power and larger memory
capacity are classified as leader robots, while the rest of the
robot are called as child robots. This arrangement seem to be
the same as the one described in [3]. The difference between
them is that in the algorithm of distributed leader-assistive
system the leader robots can use the relative measurement of
other two robot for localizing process. Thus, the child robots
are able operate beyond the sensing of the leader robots, and
still receive information from the leader robots to enhance
their pose estimation. In addition, the algorithm allows the
leader robots search for the best path to the child robots,
thereby reaching the optimal localizing accuracy and faster
respond. Another improvement of the distributed leaderassistive algorithm is that the following conditions: (i) a
given leader robot generates only a single pose
measurement; (ii) two or more leader robots do not use the
same relative measurement to synthesize the missing
relative pose measurement between leader and child robots,
and (iii) the mathematical formulation that handles the
double-counting problem will meet the requirement for

practical implementation are ensured, the problem of

overconfidence is avoided.
Although the method introduced in [8] is able to implement
effectively and eliminate some challenges in cooperative
localization, the communication cost is still high as the robot
have to communicate with other robots in team in every
time step. The algorithms in [5] and [7] have encountered
this problem. A decentralized method in [5] and distributed
approach in [7] do not require the communication between
different agents in a team in each time step. The systems
comprise all the robots are able to self-localize. The methods
introduced in [5] and [7] solve many problems of multirobot localization, but these papers have only discussed
about the issue of localization. In reality, multi-robot system
usually carry on duties in the environment where the maps
are not provided. Therefore, the paper would present the
proposed method for localizing multi-robot system in such
environments and reduce the computational cost of
cooperative SLAM method by decentralized approach in
order to make it become more competitive with other
existing method of multi-robot localization.
III.
A.

PROPOSED TECHNICAL APPROACH

Distributed and Decentralized

Simultaneous Localization and Mapping

Cooperative

The main purpose of the proposed research is to develop a

Distributed and Decentralized SLAM which was introduced
in [9]. The system model is as follows:
xi , k g( xi , k 1 , ui , k , k ), (i N )
x m , k x m , k , (m M )
y ij,,ki h(xi , k , x j , k , x m ,k , k ), (j N )( dkj ,i robs )
y im, k,i h(xi , k , xm , k , k ), (j M )(d km ,i robs )
where for timestep k :
xi , k represents the state pose of robot i
x m , k represents the state position
of the stationary landmark m
ui , k represents the odometry information of robot i
g () is the state transition function for the robots

represents the process noise

y ij,,ki represents the measurement of

landmark m with respect to robot i
h() is the measurement function

k is the measurement noise

d kj ,i is the distance between robot i and robot j

d km ,i is the distance between robot i and landmark m

B. Relative measurement using

robs is the measurement range limit

All states known, Mk is the set of landmarks:
X k {x i , k , x m, k | i N , m M k },
The states of all robots and landmarks known to robot i,
where Ni, k is the set of robots known to robot i:
X i , k {x j , k , x m , k | j N i , k , m M i , k },
The set of measurements from robot i to all robots and
landmarks within observation range:
Yi , k {y ik,,ji , y im, k,i | j N , m M , d kj ,i robs , d km ,i robs }
The set of robots that can exchange information with robot i:
Ri , k { j | j N , d kj ,i rcomm }
The centralized belief:
bel( X k ) : p ( Xk | bel( X 0),{u i ,1:k },{Yi ,1:k }, (i ))
The knowledge set, Si,k, consists of all odometry and
measurement data, as well as the previous state estimates
known to robot i at time k.
Si ,0 {bel(xi ,0 )}

The proposed algorithm is different from the method

described in [9]. This approach aims to apply the algorithm
in [8] to avoid overconfidence. That is: the relative
measurement between the two robots can only be used by
one different robot in SLAM process. For example, the
relative measurement between robot R4 and robot R3 can
only be used by R2 or R1. By this way the proposed
algorithm can handles the double-counting problem.
C. Proposed method to reduce high communication

S i,k
represent the knowledge set after state transition and
observations, but before communication is established with
any other robot:
S i,k Si , k 1 {ui , k , Yi , k }
When communication occurs between robots i and j, they
will make their knowledge sets available to each other as

jRi ,k

follows, where the

operator represents the union with
all the sets indexed by j:
Si , k S i,k S i,k
jRi ,k

From a practical and computation point of view, it is helpful

to apply the Markov property:
p ( X k | bel( X 0 ), U1:k , Y1:k ) p( X k | bel( X k 1 ), U k , Yk ),
When performing state estimation, as it limits memory and
processing requirements and allows for recursive state
estimation.
The difficulty is that in a decentralized framework, the
Markov property can only be applied once a robot obtains
sufficient information regarding other robots through
communication. Furthermore, each robot must ensure that
other robots will no longer require any of the past
information that will be discarded when applying the
Markov property. And the authors of [9] have proven that a
robots decision to invoke the Markov property does not
affect another robots ability to do the same. Furthermore, if
all robots invoke the Markov property whenever possible
based on their own knowledge, it can be sure that all robots
can still obtain a centralized-equivalent estimate.

While the method in [9] mainly focus on sparselycommunication robot network, this cause a high
communication cost when many robots operate within
sensing and communicating range of each other. In order to
solve this problem, this proposed algorithm is planning to
employ the scheduling communication in [7] to generate the
novel algorithm which may perform better than the existing
methods.
IV.

PROPOSED EVALUATION METHOD

The novel algorithm based on the existing approaches can

be tested in two stages:
Simulation: the new approach will be simulated in computer
environment, compare the results with the other methods
before implementing it in experimental scheme.
Experimental implementation: the hardware available in the
lab which are mentioned in many papers can be employed to
test the algorithm in reality.
V.

RESEARCH PLAN

The proposed research will divide into four main stages:

Stage 1: Study the algorithm introduced in [7], [8], [9].
Combine such approaches together to generate a novel
method for localization. After that simulate the result.
Stage 2: Optimize the newly generated algorithm.
Simulate the algorithm and compare the result with other
approaches. Modify the algorithm when necessary.

Stage 3: Prepare for the experimental implementation,

such as updating sensing or communicating devices
when necessary and programming embedded system.
Stage 4: Implement the algorithm and write report for it.

[6]

[7]

REFFERENCES
[1]

[2]
[3]

[4]

[5]

D. Poole and A. Mackworth, Reasoning Under Uncertainty, In

artificial Intelligence: Foundations of Computational Agents.
Cambridge University Press, 2010. Retrieved from
http://artint.info/.
S. Thrun, W. Burgard and D. Fox, "Probabilistic Robotics,"
MIT
Press,
Cambridge,
2006.
Retrieved
from
http://probabilistic-robotics.informatik.uni-freiburg.de/.
O. Silva, G. Mann and R. G. Gosine, Development of a relative
localization scheme for ground-aerial multi-robot systems, in
Proc. IEEE/RSJ Int. Conf. Intell. Robot. Syst. (IROS), Oct. 2012,
pp. 870-875.
M. W Mehrez, G. K Mann and R. G Gosine, Formation
stabilization of nonholonomic robots using nonlinear model
predictive control, in Proc. IEEE Canadian Conf. Elect.
Comput. Eng. (CCECE), Toronto, ON, Canada, May 2014, pp 16.
T. R Wanasinghe, G. K Mann and R. G Gosine, Decentralized
Cooperative Localization for Heterogeneous Multi-robot System
Using Split Covariance Intersection Filter, in Proc. IEEE
Canadian Conf. Comput. Robot. Vision. (CRV), Quebec, May
2014, pp 167-174.

[8]

[9]

[10]
[11]

[12]
[13]

O. Silva, G. K Mann and R. G Gosine, An Ultrasonic and

Vision-Based Relative Positioning Sensor for Multirobot
Localization, in Proc. IEEE Sensors Journal, March 2015, pp
1716-1726.
O. D. Silva, G. KI Mann and R. G Gosine, Efficient distributed
multi-robot localization: A target tracking inspired design, in
Proc. IEEE Int. Conf. Robot. Autom. (ICRA), May 2015, pp.
434-439.
T. R. Wanasinghe, G. K. I. Mann, and R. G. Gosine,
Distributed Leader-Assistive Localization Method for a
Heterogeneous Multirobotic System, IEEE Trans. Autom.
Scien. Eng, July 2015, pp 795-809.
K. Y. K. Leung, Cooperative localization and mapping in
sparsely-communicating robot networks, PhD thesis,
Department of Aerospace Science and Engineering, University
of Toronto, 2012.
Roumeliotis, S.I.; Bekey, George A., "Distributed multirobot
localization," in Robotics and Automation, IEEE Transactions
on , vol.18, no.5, pp.781-795, Oct 2002
Kia, S.S.; Rounds, S.F.; Martinez, S., "A centralized-equivalent
decentralized implementation of Extended Kalman Filters for
cooperative localization," in Intelligent Robots and Systems
(IROS), 2014 IEEE/RSJ International Conference on, pp.37613766, Sept. 2014.
V. Indelman, Distributed perception and estimation in multirobot systems, in Principle of Multi-robot systems Workshop
at RSS 2015.
Leonard, J.J.; Durrant-Whyte, H.F., "Mobile robot localization
by tracking geometric beacons," in Robotics and Automation,
IEEE Transactions on , vol.7, no.3, pp.376-382, Jun 1991