Sébatien Delarue, Philippe Hoppenot and Etienne Colle
IBISC, CNRS FRE 2873 - University of Evry, 40 rue du Pelvoux 91020 Evry cedex, France
Keywords: Multi-agent, arm manipulator, mobile platform.
Abstract: Assistive robotics especially mobile arm manipulator can be useful for restoring manipulation function of
disabled people in everyday life tasks. However, those systems must be designed to be reliable, fault
tolerant and easy to control. This article proposes a method based on multiagent system for controlling the
robot. This kind of distributed architecture makes possible to be fault-tolerant without specifying additional
management of faults what improves reliability. Moreover, it is also possible to add specific constraints, for
example human like behaviors in order to facilitate the use of the system by the person. The multiagent
method is easier to implement than classical robotics approaches.
ARPH project (French acronym for Robotic
Assistance to disabled People), developed in IBISC
laboratory, deals with restoring handling function for
handicapped person. It is a semi autonomous mobile
manipulator arm. Three types of control mode are
developed. In automatic one, the operator only gives
the goal and the system achieves it automatically.
However user would like to do by himself so in
manual mode, the operator controls all the degrees
of freedom of the system but the user’s workload
often is too important. The idea consists in
developing shared modes in which operator and
system share robot control. Project is divided into
two research axes: robotics for building autonomous
functions and human machine co-operation. The
paper focuses on robot control while keeping in
mind specificities due to the co-operation between
human and machine. For example industrial robot
performances such as accuracy are less necessary
than easiness of control by the user and fault
tolerance, the person being strongly tributary of the
The classical method to control a robot is to
compute mathematical static and/or dynamic models
of the robot (Yoshikawa, 1990). The approach
provides good results in known environments for
carrying out repetitive tasks. If the objective is
known in Cartesian space (
zyxp ),,( ), those
models provide speed or angular value of arm joints
so that end effector performs the task. Generally
models are computed off-line so they are not
modifiable and cannot be adapted easily to quick
changes of robot structure e.g. due to the
dysfunction of one of robot joints. It requires
addition of specific management for making the
system fault-tolerant and taking into account on-line
changes. In assistive robotics, it is an important
criterion which can be solved by exploiting the
system redundancy.
In literature, many works exploit redundancy for
other objectives, for example to keep an optimal
manipulability of the end effector (Yoshikawa,
1990), (Bayle, 2001). Some authors, (Chabane,
2005) (Yoshikawa, 1984), have proposed
manipulability measures related to the task to be
achieved. Our goal is to propose an only model
which exploits robot redundancy and able to perform
task even in case of joint default.
In addition, the model must take into account
some aspects of human machine cooperation. For
instance previous works demonstrated the interest
of giving robot human like behaviors for facilitating
the appropriation of the robot by the user
(Rybarczyk, 2002) (Fuchs, 2001).
Distributed artificial intelligence makes complex
problem resolution possible more easily than
classical approaches. Our approach is based on a
multi-agent architecture divided into two parts, a set
of agents for the arm manipulator and only one for
the mobile platform. After a bibliographical study,
we present the multi-agent architecture able to
Delarue S., Hoppenot P. and Colle E. (2007).
In Proceedings of the Fourth International Conference on Informatics in Control, Automation and Robotics, pages 292-299
DOI: 10.5220/0001625902920299
control the arm manipulator. Then, the architecture
is extended in order to control the mobile arm.
Computer science evolves from a centralized
architecture (sequential treatments) to a distributed
one (parallel treatments). So, very quickly are
appeared autonomous agents, able to achieve
individual task with no external help. MAS for Multi
Agents Systems try to solve complex problem which
cannot be solved by a unique limited-mind entity
(Ferber, 1995). An agent can be defined as a
autonomous and flexible software entity (Weiss,
1999). An agent must be able to answer environment
changes. An agent has to perceive its environment,
to treat the data and then to act. This is a sensori-
motor loop. Stimuli may come from the inside
(agent itself) or the outside (environment). Action is
performed on the agent (internal states) or on the
environment. Agent behavior is the result of
interaction between the agent and its environment.
It exists three ways to implement agent
behaviors: cognitive, reactive, hybrid. Cognitive
way divides the internal treatment into three parts
perception, planning and action. Agent must have its
own world knowledge. It is able to analyze the
situation, to anticipate and then to plan an action.
Issues of this approach are limited speed and limited
flexibility. It is also inadequate in the case of
unexpected events. Reactive agent locally perceives
its environment (and possibly its internal states) and
deduces immediately actions to be carried out only
from this source of information. This principle is
based on reflex action ((Zapata, 1992), (Wooldridge,
). Hybrid approach merges the first two agents
a basic reactive behavior with a high cognitive level.
It aims at joining reactivity with thinking and
organization abilities of cognitive agents ((Brooks,
1986), (Chaib-Draa, 2001)).
The objective of multi-agents systems (MAS) is
to bring together a set of agents and to organize
them to reach a high level goal. We can find several
kinds of MAS. There are reactive systems which
bring together agents and then try to get an emergent
behaviour that can solve a higher level problem than
each agent can do. Another type of MAS appeared
with the need of making agents communicate in
order to cooperate (Beer, 1998). Parker (Parker,
1999) uses a central machine which supervises
messages. In this case, supervisor organizes groups
of agents whose competences are different but
necessary to succeed.
Reactive MAS are often used to solve problems
with the help of limited-mind agent having poor
world knowledge and poor action ability. The
objective is to find a social emergent behaviour of an
agent society able to solve a complex problem. For
example, design of ants or bird behaviour uses this
type of approach (Drogoul, 1993). Higher level
MAS are frequently used in mobile robotics,
especially in collective robotics (Lucidarme, 2003).
Systems are based on criteria like gratification,
altruism or cooperation (Lucidarme, 2002). A
complex dialog is implemented between agents. A
high hierarchical level entity is needed to oversee
the task to achieve and to centralize decisions and
communications. The objective is for example to
coordinate several tens of mobile robots transporting
containers on quays (Alami, 1998). Some others
applications are developed for path planning,
including obstacle avoidance (potential fields)
associated with artificial life algorithms
(Tournassoud, 1992), (Mitul). In robotics an
application of MAS is the design of arm manipulator
model. We can find few approaches (Duhaut, 1999)
(Duhaut, 1993), which describe how to reach a
Cartesian position without needing arm inverse
kinematic model. The method seems to be well-
adapted to our case. Each link is implemented like
an agent. Task resolution starts with the link of end
effector. It tries to align itself with the goal and to
place end effector on the goal by uncoupling itself
from previous link. Then, next link does the same
with a new sub goal given previously.
Figure 3 shows the beginning of the recursive
algorithm applied to a 3 DOF arm. Only three steps
are shown in order to illustrate how it works. The
goal is the cross. Initial situation is first step. On step
2, the end effector link rotates virtually and
uncouples itself from the arm to reach the goal. On
step 3, second link rotates and uncouples itself from
the arm to join the end effector link. The main
characteristic of the algorithm is that end effector
link makes bigger rotation than second one and so
on. Arm unfolding is not equally distributed between
each joint.
Figure 1: D. Duhaut approach.
Another use of MAS concerns high level
complex system like fault detection (Guessoum,
1997) or data merging (vision, touch, sound ...). It
often uses neuronal networks or fuzzy rules to find
the more pertinent information. In this case the word
“agent” is usually employed in the meaning of
“expert system”.
3.1 Presentation of the Approach
In our approach, we associate an agent to each arm
joint. Each agent has the same behavior rules as the
others. A joint agent computes the position of the
end effector (
Figure 2). The objective is to move the
end effector as close as possible to the goal. The
agent behavior is described by a very simple
Agent rotates joint in one direction and computes
the new position of the end effector. If this one is
closer to the goal than the present one, it is the good
direction and rotation continues in that direction. If
the new position is not better then rotation follows
the opposite direction.
By repeating this algorithm, the end effector
moves as close as possible to the goal. In the basic
algorithm it is possible that the agent rotates in the
bad direction at the first step. We will see in the next
paragraph that it can be avoided by using arm direct
kinematic model.
Joint agent
Figure 2: Initial Agent position.
This behavior is now extended to a n-joint arm.
Each joint is controlled by an instantiation of the
reactive agent described before. The set of arm
reactive agents process data in parallel way. Agents
are autonomous and each one only tries to minimize
the distance between end effector and goal. The
algorithm of each joint is implemented in separated
processes. As said before we look for an emergent
behaviour that satisfies the global mission: the arm
end effector must reach the position of the goal.
Agent 1
ent 2
ent 3
Agent 4
Figure 3: Four agent extent.
Figure 3
shows a four agent example. As
previously explained in the one agent case, each
agent must know the position of the end effector and
the movement generated by its action. It can be done
following two ways: first one, all joint values are
recorded in real time on a blackboard readable by all
agents. So, each joint agent can calculate the
position of the end effector by using the simple
direct kinematic model of the manipulator. This
solution implies to know this model. Second way
avoids the knowledge of any model by using an
external system (video camera, GPS) that can detect
the real position of the end effector. In the first case,
as the manipulator model is known, each agent can
pre-calculate the end effector position without really
moving the joint it controls. So, after a simulation in
one direction, it can decide the right move and
rotates joint directly in the right direction. In the
second case, we don’t need models or any joint
encoders. We don’t need any communication
between agents about joint position but we must
accept that a joint may rotate in the wrong direction
at some steps of the algorithm. What is interesting in
this case is that the algorithm can be implemented
directly on any manipulator without knowing any
arm measurement.
Example of Figure 3 shows the case of a 2D
robot. The method is also efficient for a 3D system.
3.2 Results
3.2.1 Comparison with (Duhaut, 1993)
Figure 4 shows the difference between two
approaches. The first, on the left, uses our MAS
algorithm called MSMA. The second, on the right
and called MD, has been developed by Dominique
Duhaut (Duhaut, 1993). Simulation presents a five-
joint arm manipulator represented on a 2D plane.
The arm unfolding is presented in three steps.
Figure 4: Comparison MSMA and MD.
ICINCO 2007 - International Conference on Informatics in Control, Automation and Robotics
We can see that during unfolding, the MD
approach tries to unfold links starting with the end
effector one. The MSMA system unfolds all links at
the same time. So, each joint contributes to the
movement, avoiding links alignment. Moreover,
visual aspect is closer to human behavior. When a
person wants to take an object, he does not stretch
himself to the maximum.
Figure 5 shows an arm folding. Again, we can
see a more distributed movement between joints in
the case of the MSMA approach. There is no
collision between links.
Figure 5: Arm folds.
3.2.2 Behavior in Case of Broken Joint
Duhaut's method (Duhaut, 1993) does not permit to
simply take into account a joint fault so we only
present results with our method. Figure 6 shows
system behavior including two broken joints (dashed
limb and squared joint). In this simulation we can
use only the end effector position knowledge (no
kinematic model, no encoder values). In that case,
broken joint means the motor associated to it is out
of order (joint encoder out of order or not). If we use
a blackboard and the direct kinematic model, as
explained in the presentation of the approach, then,
the joint encoder must not be out of order to
guarantee a correct operation.
Here, the arm works at only 60% of its capacities.
Reachable domain is delimited by two half circles.
Dotted area represents the reachable space taking
into account the reduced capacities. A systematic
test has been realized, covering all the reachable
space: 100% of the 4592 tested positions have been
reached. The figure also shows three sample
positions. It is possible because of the redundancy of
the system.
Figure 6: Two broken joints simulation with MSMA.
3.2.3 Discussion
As agents work independently, all joints participate
to the movement inducing a natural visual aspect of
the arm for the user. For us, this property is very
important because it gives the system a human
behavior and helps human-machine cooperation.
The movement is distributed in all joints and the arm
configuration stays far from singular positions. We
can also see that there is no link collisions while the
arm folds. In the third simulation, without
integrating any fault treatment, we can see that
MSMA using external localization system is fully
fault tolerant if goal is reachable. If a blackboard is
used to store joint values in addition to the direct
kinematic model, joint encoders must not be out of
order even if motor is. In the two cases, even if a
motor is randomly moving because of noises or
control system fault, the system will reach the goal.
We can notice that MAS system does not need to be
informed of joints faults. Each software agent does
its work without taking care if it really controls or
not its associated joint.
The system to control is composed of a manipulator
arm embarked on a mobile platform. The first
objective deals with human-machine co-operation.
The idea is to give to the system behaviors inspired
from human being. For example, when a person
wants to take a book, he/she tries to keep his/her arm
not extended. If the book is too far to be taken by
arm extension, the person walks in the direction of
the book aligning the body on the hand movement
direction. The second objective is to make the
system more tolerant to some joint fault by
exploiting redundancy without using specifics fault
treatments. Indeed, in assistive field, it is important
to maintain a good quality of service.
4.1 Mobile Platform Agent
The control of the mobile platform uses an agent
which controls angular and linear speeds. The
mobile platform agent is more complex than the arm
ones. The agent used for the mobile platform is an
hybrid one. Its cognitive capacities give the
possibility to add some interesting behaviors.
Firstly, the mobile platform has to move forward
because sensors for obstacle avoidance are located
on the front side of the system.
The second implemented behavior is to align
direction of displacement of the mobile platform
with the manipulator arm. Indeed, when a person
wishes to catch an object and must move to do it,
he/she generally tights the arm forward in the
direction of the movement of the body.
The third behavior concerns deadlock. In certain
cases, both mobile platform and arm shoulder joint
rotate at the same speed, but one on the right and the
other on the left direction. In that case, the gripper
does not converge to the objective. The agent of the
platform is able to detect this kind of situation. It
introduces a waiting cycle by leaving to the arm the
priority for reaching the goal. Once this cycle ends,
the platform agent tries again to align itself with the
arm if it is possible.
Reactive behavior is the same one as for the arm
agents. Mobile platform agent tries to minimize
distance between the end effector and the goal.
Mobile platform agent works independently from
the arm ones.
4.2 Results
We now simulate the whole system algorithm. First
we compare it with a classical mathematical method
using manipulability constraints. Secondly, we
simulate faults on some joints.
4.2.1 System and Protocol
The system is a 3D mobile arm. It is composed of a
mobile platform equipped with two driving wheels
and a free wheel to ensure stability. A 6 DOF
manipulator arm is fixed on the mobile platform. In
the following simulations, the end effector imposed
task consists in following a straight line with a
constant speed, in the upper-right direction of the
mobile platform.
Figure 7 describes the system and
the associated mathematical frame. The
displacement is perpendicular to the initial
orientation of the platform. The displacement plan is
formed by x and y axis. Initial conditions are
showed in
Table 1.
Figure 7: System description.
Table 1: Initial simulation conditions.
Object Initial value
Platform position (0,0,15) meters
Manus Joint 1 270 degrees
Manus Joint 2 120 degrees
Manus Joint 3 -125 degrees
End effector position (18,79,-20) meters
Simulation steps 400
Sampling rate 60 ms
Shift wanted for each step (0.42 , 0.12 , 0) cm
Total duration 24 s
On the first simulation, objectives are the
following ones:
- End effector must follow the desired trajectory
with good accuracy
- Mobile platform must move forward because of
front side sensors belt
- Mobile platform has to align itself with the arm
- Arm has to avoid its extended configurations.
We compare our approach (MSMA) using
agents introduced above with the MIM one which
uses manipulability criterion (Chabane, 2005).
Secondly, we check fault tolerance ability of
MSMA approach. So, we simulate a breakdown of
the shoulder and check if the whole system
redundancy (generated by the mobile platform)
permits to the system to reach the goal. We also
simulate a breakdown of joint 2 to check the whole
system behavior.
4.2.2 Results
Her, we do not show end effector trajectory. With
both approaches, the move is correct with a good
accuracy (less than 3mm of difference between the
real path and the desired one). There is no notable
ICINCO 2007 - International Conference on Informatics in Control, Automation and Robotics
difference between them. Accuracy is linked to
simulation step duration. We choose a high one of
60 ms because our real system has a 60 ms
command loop.
Figure 8 shows platform trajectory and
orientation. We can see two very different
performances. MSMA works well. It always goes
forward keeping obstacle avoidance sensors in front.
It first turns slowly on the right, and then goes
straight until the end and aligns itself with the arm
orientation. We see a graining point in MIM
simulation and the platform ends the task moving
During the move and with both models, the arm
is never bended. Angle between joint 2 and joint 3
stabilizes to a 70 degrees value which is far from the
0 degrees singular value of the arm.
Figure 8: MIM and MSMA platform moves.
We now show how the system reacts in fault
cases. MIM is not designed to be fault tolerant so its
behavior is not interesting here. It is why next
simulations show only MSMA model in three
samples faults situations:
Arm breakdown joint 1 (shoulder) at 60°
Arm breakdown joint 1 (shoulder) at 30°
Arm breakdown joint 2 at 120°
These values are chosen as examples to illustrate
motor joint breakdown fault tolerance. We assume
that if goal is reachable, fault tolerance is effective
in all the cases. Breakdown occurs at time 0. Figure
9 shows end effector trajectory on x axis. We can
see that MSMA satisfy objective even in fault
situations. Results are the same on y and z axis. In
the first two situations, redundancy between arm
joint 1 and the mobile platform rotating movement is
exploited. As end effector follows the wanted path,
we can say that the platform agent works well when
the arm shoulder is broken. In that case, mobile
platform moves forward but does not align with the
arm orientation. Indeed, this alignment is performed
by the redundancy between arm joint 1 and the
mobile platform rotating movement and, in that case,
joint 1 is broken. Moreover, the arm is not tightened
and the angle between joint 2 and 3 is stabilized to
70 degrees. In the third situation, redundancy
between joint 1 and joint 3 is exploited. Both of
them permit a vertical move (y axis). Once again the
end effector follows the trajectory correctly. In that
case, the mobile platform moves forward and aligns
with the arm orientation. We can still notice that the
arm is not tightened and that its posture remains far
from singular position.
Figure 9: End effector trajectory on axis x.
4.2.3 Discussion
On the first simulation, we can see that the classical
model used implies a more complex trajectory for
the mobile platform on the whole system simulation.
There is a graining point making the robot blind (it
can not avoid obstacles anymore because ultrasonic
belt is in front side). It’s due to the manipulability
criterion used in this simulation. Our model shows a
better behavior, closer to a human one. The platform
goes forward and aligns itself on the arm orientation
inducing a more assimilable move for the user. The
use of independent agent helps to considerate
directly human behavior in algorithm. Also, the
system is fault tolerant as shown in simulation. The
end effector desired trajectory is reached even if arm
joint 1 or 2 are broken. This ability is due to multi
agent architecture which is able to run even if a
component is defective.
First, with our approach, we do not cut out the final
objective in sub goals which each agent would have
to reach. Each one has a local work to do
independently from others. We do not organize any
co-operation. Here, we speak about emergent
behavior. Indeed, one agent can not reach the goal
alone. It needs others to achieve the task but it does
not know it. This kind of system provides very good
result concerning fault-tolerance.
Secondly, this approach induces a goal for each
agent. It is then possible to influence behavior of
some agents. Then we can easily adjust or add
behaviors to facilitate man machine co-operation.
That is the case for example with platform alignment
on the direction of the end effector in the same way
than alignment of the human body on its hand
direction. That leads to an easier assimilation of the
system by the user. Indeed, with a classical model,
to add a behavior requires the integration of
constraints in the global model itself, which is not
easy with this kind of system.
Our approach has also its limits. If we integrate
many behaviors, it is possible for the system to loose
its wanted emergent behavior. The added constraints
could be in conflict with the initial objective which
is to reach the goal. The system then enters in non
convergence cases. At the same time, we can loose
fault tolerance ability. To avoid this kind of errors, it
is first necessary for agents to keep autonomy
compared to the goal they have to achieve and thus
to be as simple as possible. Secondly, we have to
include priorities in the local objectives of each
agent. Reaching the goal has a high priority, going
forward has a smaller one. Aligning platform on arm
orientation has a very small one. Thirdly, we have to
implement deadlocking treatment by introducing
delays in specific situations. In our system, these
particular treatments are implemented in the agent of
the mobile platform. We do not plan any problem
resolution between agents. In our approach, we keep
simple algorithm to avoid high hierarchical
management. Indeed, high hierarchical management
could then be compared with a system using
mathematical models and including a fault treatment
supervisor to switch between them.
Our MAS system gives good results in relation to
human behavior. Objects can be caught with an
easily assimilable movement for the user (forward
move, alignment of the platform with arm
orientation, simple trajectory with no turnaround, no
bended arm). Accuracy is similar to classical
method. It is fault tolerant without integrating any
specific treatment. It makes easier the integration of
special human driving mode. There is no need to
compute mathematical model and especially the
inverse model. Our MAS system algorithm is easy to
implant and need only some geometric formulas and
thus very little computing power. It is a real time
one. We now have to implement algorithm on our
mobile platform and create scenarios of
displacement to judge the relevance of our algorithm
on a real robot. The first tests on the real system give
encouraging results. Simulations shown in this
article give an idea of the mobile manipulator
behavior with straight lines trajectory. These kinds
of trajectories are usually used by people driving
ARPH with HMI actually developed. So we think
that MAS system can directly replace the Cartesian
internal manipulator mode actually used. Actual
HMI allows user to control the platform and the arm
manipulator separately (two sets of buttons) as
shown on Figure 8.
Figure 10: The actual HMI two sets of buttons.
With MAS, it’s possible to control the whole
system with only one set of buttons by driving only
the grip, thus limiting the difficulty of seizure
operation by the user.
We also want to improve object grasping. One
possible way is to integrate a neuronal network that
could help the system to have a better posture (eg:
catching an object by the left if user is a left hand
writer...). This improvement should lead us to
manage with agents not only for the mobile arm but
also for the orientation of the grip. Moreover, ARPH
mobile platform has an ultrasonic sensors belt for
obstacle avoidance when moving in a room that has
to be integrated with MAS system to have a fully
operational system.
Yoshikawa, T.: Foundation of robotics: “Analysis and
control”, The MIT Press, 1990
Bayle, B., Fourquet, J.Y. and Renaud, M.:
“Manipulability Analysis for Mobile Manipulators”,
ICRA’2001, Séoul, South Korea, May 2001, pp. 1251-
Khiar Nait Chabane; Philippe Hoppenot; Etienne Colle,
“Mobile Arm for Disabled People Assistance
Manipulability Analysis”, International Computer
Systems and Information Technology Conference,
Algeria, July 2005
Yoshikawa, T.: “Analysis and control of Robot
manipulators with redundancy”, In M. Brady & R.
ICINCO 2007 - International Conference on Informatics in Control, Automation and Robotics
Paul. Robotics Research: The First International
Symposium, MIT Press, 1984, pp 735-747.
Yves Rybarczyk, Etienne Colle, Philippe Hoppenot:
"Contribution of neuroscience to the teleoperation of
rehabilitation robot", SMC'2002, Hammanet, Tunisia,
6-9 october
Fuchs, J. Moreau, G et Papin, “Le traité de la réalité
virtuelle“, Nantes : Les Presses de l’Ecole des Mines,
Jacques Ferber, "Les systèmes Multi agents Vers une
intelligence collective", IIA Informatique Intelligence
Artificielle InterEditions, september 1995
Weiss G., "A Modern Approach to Distributed Artificial
Intelligence", The MIT Press, Cambridge,
Massachusetts , 1999
R. Zapata, P. Lepinay, C. Novalés and P. Déplanques,
"Reactive behaviors of fast mobile robots in
unstructured environments : sensored based control
and neural networks." SAB, December 1992.
Wooldridge M., Weiss G., "Intelligent Agents" Multiagent
systems, - Ed.: MIT Press - 1999.
R. A. Brooks, "A robust layered control system for a
mobile robot", IEEE Journal of Robotics and
Automation, p. 14-23, 1986.
Chaib-Draa B., Jarras I., Moulin B., "Systèmes multi-
agents : principes généraux et applications ,Principes
et architectures des systèmes multi-agents, J.-P. Briot
et Y.Demazeau (Eds.) Hermès - 2001.
Beer M., d’Inverno M., Luck M., Jennings N., Preist C.,
Schroeder M., "Negotiation in Multi-Agent Systems",
Workshop of the UK Special Interest Group on Multi-
Agent Systems (UKMAS’98) - 1998
L. Parker, "Adaptative heterogeneous multi-robot teams",
Neurocomputing, special issue of NEURAP’98:
Neural networks and their applications, 28, p.75-92,
Drogoul A, "De la simulation multi-agent à la résolution
collective de problèmes. Une étude de l’émergence de
structures d’organisation dans les systèmes multi-
agents.", Thèse de doctorat, Université Paris 6. - 1993
P. Lucidarme and A. Liégeois, "Apprentissage de
comportements réactifs pour des ensembles de robots
mobiles",JJCR17, Paris, pp. 33-39, IEEE best paper
award - 2003
P. Lucidarme, O. Simonin et A. Liégeois, "Implementation
and evaluation of a satisfaction/altruisme-based
architecture for multi-robot systems", proc. ICRA'02,
Washington D.C., pp. 1007-1012 - 2002
R. Alami, S. Fleury, M. Herrb, F. Ingrand and F. Robert
"Multi-robot cooperation in the Martha project",
IEEE Robotics and AutomationMagazine, 5(1), p.36-
47, 1998.
Pierre Tournassoud, "Planification et contrôle en
robotique Application aux robots mobiles et
manipulateurs", Edition Hermes Traité des nouvelles
Technologies série robotique – 1992
MITUL SAHA, "Manipulator Planning and Obstacle
Avoidance Using Genetic Algorithm",
Dominique Duhaut, "Distributed Algorithm For High
Control Robotics Structures", International
Conference On Artificial Intelligence – 1999 –
Volume 1 p 45 à 50
Dominique Duhaut, "Using multi agent approach solve
inverse kinematic problem", proceeding 1993
Z. Guessoum, "A Hybrid Agent Model: a Reactive and
Cognitive Behavior", ISAD97, IEEE, Berlin,
Germany, pp. 25-32 - April 1997.