JOSEP LLUIS DE LA ROSA peplluis@eia.udg.es
ALBERT OLLER aoller@etse.urv.es
Universitat de Girona
Universitat Rovira i Virgili

Abstract: Multi-Agent decision-making structures are expected to be extensively applied in complex industrial systems. Multi-robotic soccer competitions are a forum where rules and constraints are good to develop for MIROSOT, and test them. This paper describes an implementation of such a robotic team consisting of three micro-robots. It shows an agent-based decision-making method that takes into account an index of difficulty to execute collision-free path obtained by a path-planning method specially devised for this domain. Some real situations are used to explain how robots behave, and which are their movements through the playground. Then, some agents persistence problems are analyzed, and simulation results are shown.

Keywords : Decision-making, Multi-Agent Systems, Co-operation, Multi-robots, MIROSOT.


Decision-making structures are expected to be applied in complex industrial systems (namely distributed systems like energy or material distribution networks, big production plants with several stations and dynamical or discrete processes with huge number of variables)

These systems are so-called Complex Systems (COSY) because all of them have some common properties: complexity, dynamism, uncertainty, and goal variability. COSY can be physically distributed, i.e., resources, data bases, plants, and so on.

To facilitate the automatic control community to develop agents in control systems analysis, design, and implementation, this work proposes to incorporate agents’ capabilities and facilities in Computer Aided Control Systems Design frameworks (CACSD). These facilities were initially introduced in [ 1] by means of an Agent-Oriented Language (AOL) called AGENT0[ 2] that consists of a language to program agents by means of belief, goals, commitments and capabilities.

This system is developed to do some research on physical agents design, and practical results are obtained from MIROSOT’97 [ 3] competition. Because of FIRA [ 4] rules, robots are limited to the size 7,5´ 7,5´ 7,5cm. and must run autonomously through a 130´ 90 cm. sized ground. So far, teams can use a centralized vision system to provide position and orientation of the robots so as ball position. Does this domain implies Agent Distributed Artificial Intelligence? This domain possesses many characteristics given that it is a real-world system, with multiple collaborating and competing robots. Here follows some properties [ 5] :

Complexity. Playmates do not possess perfect information about the game field, provided by the vision system.

Dynamism. Multiple playmates or opponents robots can independently execute actions in this environment.

Uncertainty. A real-time system with many vehicles ensures that even identical start states lead to very different scenario configurations and outcomes.

Goal variability. Playmates are constantly moving, then an individual robot at any one time is executing individual actions. When team plan changes, i.e., attack instead of defend, individual goals change instantaneously.

This properties fulfil in the COSY. This micro-robotic system is designed to get up to 0.6 m/s controlled motion. Work went on specializing the approach of automatic control community by introducing the classical structure behavior- supervision- control algorithms [ 6][ 7] . Thus, both the multi color based-vision system and the micro-robot control structure are specified in this paper.

This work focuses on how co-operation is implemented in this dynamic and real-world (real-time) domain. This co-operation among robots is introduced at the behavior (agent) level, that is also responsible for the behavior of each robot by exchanging messages to the supervision level (see "Fig.K").

Finally, this work has developed the soccer system that contains the following new features:

1) Explicit partition of the behavior-supervision-control levels, developing mainly the behavior-supervision levels.

2) Exchange of information between behavior « supervisor levels. The supervisor (developed with toolboxes with MATLAB/SIMULINK (M/S) [ 8] ) executes targets proposed by the behavior level, and the behavior level obtains suggestions about difficulty indexes in possible targets to feed its co-operative behavior.

3) The exchange of fuzzy evaluations in the behavior « supervisor communication is included. The fuzzy capabilities in the AOL of the behavior can use now fuzzy evaluations of difficulty indexes in possible targets.


A. What Agents Are?

The agent A beliefs in good_weather that agent B has told him. Before this first agent introduces this belief in his knowledge base together with other evidences he will criticise this information, that could be called modification of beliefs. The process of critics or review could work with three hypotheses:

  • Agent B wants that A goes out with him on excursion, and therefore is perfectly possible that he's trying to lie him. Perhaps A will not belief that good_weather because B isn't reliable.
  • Agent B, is an unreliable system because its reasoning or perception isn't good. For instance, that continuously says good_weather regardless of the current weather. In this case, the certitude coefficients that could emit aren't neither reliable.
  • Agent A beliefs whatever B informs (classical approach in systems exchange of information).
  • B. What Physical Agents Are?

    Definition 1: Software agents. This term denotes a software-based computer system that has several properties [19] as autonomy, social ability, reactivity, pro-activeness, mobility, rationality, etc.

    Physical Agents are software agents that contain the N/S (Numerical/Symbolical) and S/N (Symbolical / Numerical) interface that is typical of real systems, which according to [1] and [8] are constrained by imprecision, uncertainty, changing through time, and others.

    One typical implementation of physical agents (but not the unique) is mobile robots, that in current research are progressively more and more autonomous and co-operative. The traditional AI has focused on symbolic paradigms (toy problems) and has not expended time on real applications. On the other hand, robotics has focused on design and construction of hardware and its control.

    For solving current problems in autonomous robots, traditional AI has evolved into perception based and multi-agent approaches. The conjunction of AI and robotics in autonomous mobile robots that solve in an emergent way, complex problems by Cupertino is important, especially when the environment continuously changes because of the movement of the physical agents.

    Having a "physical body" according to [18] could be summarised as:

  • Sensorial and action capacities are closely related.

    Josep Lluís de la Rosa es Doctor en Ciencias por la Universidad Autónoma de Barcelona desde 1993 http://eia.udg.es/~peplluis. Tras consecutivas estancias pre y postdoctorales en Francia, es el director del grupo de investigación eXiT, Ingeniería de Control y Sistemas Inteligentes, desde el 1994, http://eia.udg.es/exit. Sus trabajos más conocidos son los agentes físicos y especialmente los relacionados con los desarrollos con robots futbolistas, http://rogiteam.udg.es

    Albert Oller: Físico por la Universidad Autónoma de Barcelona. Actualmente es profesor asociado. La investigación se centra en el ámbito del diseño de agentes físicos dinámicos.