Use of Multi Agent Systems for Mobile Robotics Control

J.Y. Tigli and M.C. Thomas

Universitd de Nice - Sophia Antipolis Laboratoire 13s-CNRS URA 1376 Bat 4, 250 Rue Albert Einstein , Sophia Antipolis 06560 Valbonne, FRANCE email : tigliQmim,osa.unice.fr

Abstract Multi Agent Systems in Mobile Robotics are generally use to represent Multi Robot Systems and to study the interaction between robots to achieve a global task. But mobUe robot agent8 are designed in different waya, with a reactive level and sometimes a cognitive level to provide the behavior of the agent with a deliberative decidon making. After a brief survey of the use of Multi Agent Systems in Mobile Robotics, we explore the pouibility to design a mobile robot agent using an other Multi Agent System based on selfinterested autonomous entities of the mobile robot. I INTRODUCTION

A.

Multi Robots Systems

The complexity of systems due to multiple autonomous entities like mobile robots, makes Multi Agent System an attractive concept. Generally, for lack of one finite model only of such complex systems (in Robotics, Telerobotics, Process Control ..etc.), Multi Agent concept is useful. They provide applications with distributed decision making and robustness face to unexpected events (for example agent failures). In this paper we present a brief survey on the use of Multi Agent Systems in Mobile Robotics, generally to coordinate autonomous robot societies. Then we present different approaches to design such mobile robot agents, especially in their reactive and deliberative aspects. In the second part, we explore the possibility to use a Multi Agent System to design such agents, using a simple example implementing agents as self-interested autonomous physical parts of a mobile robot. Finally, we propose a Multi Agent System to design reactive level of an intelligent control architecture of a mobile robot, implementing agents as behavioral entities .
11. A MOBILE ROBOT IS AN AGENT

Generally in the literature, a whole autonomous mobile robot is considered as an agent, and the Multi Agent concept in this case is used to study interaction mechanisms between these entities to result in a selforganization of the set of agents. In this wa , Fukuda proposes the CEBOT system [lo] to study d e self-organizing behavior of heterogeneous robotic agents, called Cellular Robots. For example, in [13], they use Mobile cells to carry the same object coordinately. In this approach each cell must obey simple controI laws. The resulting distributed decision making of this kind of multi agent robotic system is then studied as the conditions on the different parameters of these control laws to keep the t e tal system stable. The efficiency of such an approach is evaluated according to the number of agents used to achieve the task. Other works were developed in the same way, using reactive agents to perform a task r in dangerous o inaccessible environments for human beings. Brooks [7] at MIT uses Subsumption-based multi agent teams a8 mobile robot builders for a lunar base construction. Likewise Steels uses micro robot societies to collect some sampIes in other planets and studies self-organization between reactive agents [191, Arkin evaluates the impact of the absence of interagent communications or of simple communications between agents to achieve a given task 21. But in fact the eEciency of a Multi Agent System epends on the kind of agent used and how we design it.

B.

Design of a Mobrle Robot Agent

Research is currently being conducted on the use of multiple robots to solve problems in a changin workspace and more particularly to achieve a globa task.

If

Autonomous mobile robots can be strictly reactive or CO nitive reactive agents, according to the use or not o a deli erative level in their decision making. Of course in the case of purely reactive agents, when Intelligence is determined by the dynamics of interaction with the world [6], a self-organization of non-intelligent mobile robots about a same goal is more spectacular. Nevertheless such an approach can present some short comings due to the lack of flexibility to take advantage of a priori knowledge [ll], when the robots must perform a different task. In the

a 6

0-7803-2129-4/94 $3.00 Q 1994 IEEE

Authorized licensed use limited to: UR Sophia Antipolis. Downloaded on December 8, 2009 at 13:50 from IEEE Xplore. Restrictions apply.

robotica field, reactive agents have gained popularity. They are characterized by a reactive low-level control, to ensure the faatest feedback loo through the real world (ex : clam of behaviors [4J motor schemas reflexive behaviors [ l q , activities [9] ..etc.). rooks uses subsumption mechanism to coordinate these different loop : an arbitration mechanism based on competence levels of the agent, allows to inhibit outputs of lower-levels of the agent. The lack of flexibility of such static priorities between competence levela lead him to introduce the activation level concept [14] which can be modified according to spread of m a sage between reactive layers (hormones [SI). Kaelbling [12] propose an approach baaed on the situated-automata concept to apply relevant actions on the environment according to the sensory information and the current state of the system. In the stme way [9] lay strees on need of the global understanding of the situation of the mobile robot face to its environment. So they introduce a high-level controller, blackboard baeed, to elect the relevant activity of the system from the overall sensory informations collected on the blackboard.

k],

p8j,

whole mobile robot. Only a few papers deal with the posaibility to call an Agent a component of a mobile robot. Here we are going to preeent a simple example of such an approach using a Multi Agent System to design an intelligent f control architecture o a Mobile robot. Here we consider an agent as a self-interested autonomous physical entity. s Thie example i stron ly influenced by the Braitenbergs Vehicles [3]. In fact in this example we choose f some couplea o sensora/actuators to describe our agents.

A.

The agent
wheel
\

lx0-Q sensor

Aeent

Motions

Other approaches try to combine deliberative and reactive aspects in the design of a mobile robot agent, even if the word Agent is not always used to represent the mobile robot as an autonomous self-interested entity. Arkin uses a combination mechanism (a kind o otential field method) to combine the effects of its f di erent motor schemas. Nevertheless he introduces world knowledges to improve the system strate configuring behawors towards a more flexible f&nerbased system [l]. Simmons proposes an architecture to coordinate the agents planning , perception and action, taking into account the limited physical and ak computational resources usin his T s Control Architecture [MI. Muller [16] deines his distributed autonomous system architecture using two blackboard baaed levels. The first one, mana ea the reactiveness of the mobile robot and select atehavior according to the current sensory informations and the state of each behavior. The second level is the cognitive layer of the system. the corresponding blackboard contains the topological graph of the environment updated by the learner module and provides required informations for the localizer and planner modules. This cognitive level selects a preferred behavior at the reactive level. In fact Multi Agent Systems in the mobile robotics field Beems to be the study of the interaction between autonomous mobile robots. The main application is then to coordinate such a ents into a same overall goal. On the other hand t%e design of mobile robot agent can use various approaches and concepts, can be limited to a reactive level or can integrate cognit ive layers.

Figure 1: Description of a wheel Agent (Vehicle 1) In the case of the Braintenbergs Vehicle, an agent consists of one or various proximity sensors (perception capabilities), a little bit of circuitry (decision capabilities), and a motor wheel (action capabilities). In the case of the Vehicle 1, the agent can only o straight forward or backward always in the same rection. When it finds an obstacle, he kee s a constant distance between them. The behavior ofthe agent is then represented by the equation :

&-

V;. = Vmat - K-yj (1) where K is the velocity of the wheel of the a ent (action capability), and gj is the measurement from the proximity sensor (perception capability) (figure 2). Of
Y :

t
I

distance to the obstack

Figure 2: Distance to the obstacle and measurement course such an agent is very limited for its motions and only one agent isnt very interesting for an application even if it is autonomous and self-interested.

111. CAN A MOBILE ROBOT BE A MULTI AGENT SYSTEM ?

General1 , the design of an intelligent control architecture dbesnt use the Multi Agent concept, probably because the autonomy of each considered entity is riot so clear aa the autonomy and the self-interest, of a

B.

The corresponding Multi Agent System

A mobile robot with two differential wheels is an example in this case of Multi Agent System using two

589

of these agents. Each agent can always select its velocity according to

Figure 3: Description of a Mobile Robot its measurement, but there are constraints that we can consider as interaction conditions between both agent as the mechanical link between both wheels Both wheels are fixed on the m e axis interaction can be expressed with t e following equaA p8Tt O this tion : 2.r.e - vi v = 0 a (2) In this way each agent has its own behavior and for example, we can exprass the emergent rotation of the - -mobile robot face to the first agent decision making as Figure 4: Description of a behavioral Agent perceptual level is integrated in the corresponding basic reactive behavior of the agent. The behavioral part represents the basic behaviors supplied by the agent to the mobile robot. This is the only way for the agent to act in the real environment. The decision making part evaluates the conditions to activate or disactivate the agent according to the current situation perceived by the agent and its opportunistic strategy.
2) The corresponding Multi Agent System. :

1;

where the first brackets represent the own decision making of the first agent involving its perception y1 and the second, the influence of the other agent. The resulting behavior of the mobile robot corresponds to one Vehicle 2 in the Braintenberg's classification. The mobile robot searches for obstacles and keeps face to the first encountered with the desired distance. This example is only a simple one to explain how we could comider a physical autonomous entity of a mobile robot as an agent and then the mobile robot itself as a Multi Agent System. Then we are going to present our Multi Agent System approach to design the reactive level of a mobile robot architecture. In this caae the agent is a virtual entity and a component of the mobile robot reactive behavior.
1V WHEN AN AGENT IS A BEHAVIORAL ENTITY

In oiir approach, the agent is a virtual entity in the mobile robot design. In fact each agent is a behavioral entity for the mobile robot and gives it a basic behavior face to its environment. Each basic behavior is autonomous because they can work alone, and self-interested because it has its own goal ( to avoid obstacle, to follow the environment on the left ...etc).

1) T h e agent :
The agent consists of three parts (figure 4) : The perceptual part is divided into two levels. The first level generates a symbolic representation of the current situation of the mobile robot needed by the decision making part. The second

Like we have seen in section 2 of this paper, we can use different interaction principles like subsumption mechanism, situated approach, combination mechanism ...etc. In the case of an arbitration mechanism we must verify that only one behavior is running at a time to avoid unmanaged conflicts. In the case of a combination mechanism between agents this problem is avoided but the system is harder to tune (The emergent behavior is not a basic one but a fusion of different behaviors). In our agent model, our approach is a situation oriented one. Each agent is activated accordinq to its own activation conditions on the current situation. To avoid conflicts we can add priorities between agents to choose the best one to activate. In our experiment with the mobile robot Khepera [15 (figure 5 ) , we use three agents and so three kinds of asic behaviors. Other behaviors are easy to build using neural networks, control laws ...etc. The behavior of the first agent is the Braitenber s Vehicle 3 to avoid obstacles. This behavior uses t i e six infrared sensors on the front of the mobile robot.

K = Vmaz i=4 3

Ki.Yi

(4)

VZ = Vmaz i=l

Ki-Yi

(5)

590

V.

Figure 5: Description of Khepera mobile robot

Ki ate the different coefficients to tune the obstacle avoidance. The behavior of the aecond a ent allows to follow the environment on the left side. t h i s is a control law that servea the distance between the sensor 1 and the obal stacle on the left (claesically a w l ) and the distance between the saw 3 and the obstacle on the left front of the mobile robot. In fact the corresponding behavior is following the obstacles on the left and slowing down wheen it meets an obstacle.
(6) (7) K1 and K are chosen to stabilize the system. Of 2 course this a ent is more precise but less robust bec a w of the imits of the sensors (figure 2) and the imperfect model of the environment. Fortunately, we can choose when such an agent is really efficient and can be activated according to the representation of the current situation and the activation conditions of the agent. The third agent looks like the previous one, but to follow the environment on the right of the mobile robot and to slow down when an obstacle appears in its right front.

v = Kl-(yl - yld) = K2.(92 - Bd)

Figure 6: Reactive Navigation of Khepera Thus we obtain an overall robust behavior (figure 6) of the mobile robot as the situated interaction between the corresponding basic behaviors of each agents.
IV.CONCLUSION

We can have a complex representation of the current situation of the mobile robot according to the sophistication of the first level of the perceptual part in our agents to choose the most relevant basic behavior at each time. In fact in our experiment we choose anly a simple representation of the current situation of the mobile robot according to the symbolic information of the proximity sensors : for each sensor, is it stimulated or not ? Thus the first level of the perceptual part put a threshold on the sensor measurement to consider it as stimulated or not and so to give the corresponding binary data for each infra-red sensor. In this way the symbolic representation of the situation consists of eight binary data :
PFl(stimulated, not-stimulated) IF2( st imulated, not-st imulated) I F 3 (at imulat ed , not-st imulated)

We have presented different uses of Multi Agent Systems and more particularly in the mobile robotics field. Generally, a mobile robot may be considered as an agent and a Multi Robot System studied as a Multi Agent System. We have seen that a lot of different approachea can be used to design a mobile robot as an agent and it seems very difficult to find a common representation to describe and compare these different approaches. In the second part of this paper we explore the possibility to describe such systems as Multi Agent Systems by decoupling a mobile robot into different virtual or physical autonomous entities. In fact the main idea of this paper is to ask both questions : is an agent an autonomous module ?, and can an agent be designed as a Multi Agent System itself?
REFERENCES

~..

R.C. Arkin. Motor schema baaed navigation for a mobile robot : An Approach to Programming by Behavior. In Proceedings of ZEEE International Conference on Robotics and Automation,

Thus activation conditions for each agent are :

pages 264-271, Raleigh, North Carolina, April 1987. IEEE computer society press.

59 1

Authorized licensed use limited to: UR Sophia Antipolis. Downloaded on December 8, 2009 at 13:50 from IEEE Xplore. Restrictions apply.

[2] R.C. Arkin, T. Balch, and E. Nitz. Communication of Behavioral State in Multi-agent Retrieval Tasks. In Proceedings ofthe ZEEE Zntema-

[15] F. Mondada, E. Franzi, and P. Ienne.

Mo-

tional Confemme on Robotics and Automation, volume 3, pa es 588-594,Atlanta, Georgia USA, May 1993.I#EE Computer Society Press.

bile robot miniaturisation : A tool for investip tion in control algorithms. In Pmeedings of the third International Symposium on Experimental Robotics, Kyoto, Japan, October 1993.

[3] V. Braitenberg. Vehacles. MIT Press, 1986.

[4]R.A. Brooks. A robust layered control system for a mobile robot. IEEE Journal of Robotics and Automation, 2:14-23, April 1986. [5]R.A. Brooks. Inte rated a stems based on behaviors. SZGART Butetin, 274):46-50, August 1991. [6] R.A. Brooks. Intelligence without r e m n . In Prep a d for Computers and Thought, International Joint Conference on Artificial Intelligence - ZJCAZ91, 1991.
[7] R.A. Brooks, P.M a a , M. Mataric, and G. More.

(161 J.-P. Muller. A distributed autonomous system architecture. In Proceedings of International Confemnce on Systems, Man and Cybemetacs, volume 2,pages 623-625,Le Touquet, October 1993.
IEEE.
[17] D. W. Payton. An architecture for reflexive autonomous vehicle control. In Proceeding8 of ZEEE

Robotics and Automation, volume 3, pages 18381845,San Francisco, California USA, April 1986. IEEE computer society press.

[18]R. Simmons. Coordinating planning, perception, and action for mobile robots. SZGART Bulletin, 2( 4): 156-1 59,August 1991 . [19] L. Steels. Cooperation between diertibuted agents
through self-organieation. In

Lunar base construction robots. In Z E ZnEE ternational Workshop on Intelligent Robots and Systems (IROS90), pages 389-392, Tsuchiura, Japan, 1990.

Y.Demazeau and

J.-P. Muller, editors, Pwxeedings of Decentmlized Artificial Intelligence, North-Holland, 1990.

[8]J.H. Connell. Creature design with the subsumption architecture. In International Joint Conference on Artificial Zntelltgence - IJCAI 87, Milan, Italy, August 1987. [9] R.E. Fayek, R. Liscano, and G.M.Karam. A
system architecture for a mobile robot based on activities and blackboard control unit. In Pmceedings of IEEE Robotics and Automation, voleri, ume 2, pages 267-274, Atlanta, G o g a USA, May 1993.IEEE computer society press.

[lo] T. Fukuda, S. Nakagaya, Y. Kawauchi, and M. Buss. Structure decision method for self organisin robots baaed on cell structurekcebot. In

Procedngs of ZEEE Robotics and Automation, volume 2, page% 696-700, Scottadale, Arizona, USA, May 1989.IEEE computer society press.

[ll] E. Gat. Integrating reaction and planning in a hetero eneous asynchronous architecture for mobile rofot navigation. SZGART Bulletin, 2(4):7074,August 1991. [12]L. P. Kaelbling. G o d as parallel rogram specifications. In Pmeedings of AAAf88, volume 1, pages 60-64, Saint Paul, Minnesota, USA, August 1988.
[13] Y.Kawauchi, M. Inaba, and T. Fukuda. A prinf ciple o distributed decision making of cellular robotic system (cebot). In Proceedings of ZEEE Robotics and Automation, volume 3, p es 833838, Atlanta, Georgia, USA, May 198. IEEE computer society press.

1 4 P. Maes. The dvnamics of action selection. In 11 Proceedings of ZjCAZ89, pages 991-997,Detroit, ME,USA, 1989.
L--1
~

592

Authorized licensed use limited to: UR Sophia Antipolis. Downloaded on December 8, 2009 at 13:50 from IEEE Xplore. Restrictions apply.