Hierarchical Fuzzy Inference System for Robotic Pursuit Evasion Task Daniel Hládek * , Ján Vaščák * , Peter Sinčák * • Center for Intelligent Technologies, Department of Cybernetics and Artificial Intelligence, Technical University Kosice, Slovak Republic daniel.hladek@tuke.sk Abstract— We propose hierarchical multi agent control system based on rule based fuzzy system for pursuit-evasion task and state a new representation of this type of game that is based on fuzzy logic. This approach enables improvement of the rule base under uncertain conditions and can process a priori inserted expert knowledge. Example application domain includes reckon and guard robots, research space probes, coordination of multiple mine sweeping devices or autonomous rescue teams.. I. INTRODUCTION Cooperation of multiple robot formation executing common task might be very difficult. These tasks require ability to deal with unknown and uncertain situation by single entity and also take into account success of the whole team. We propose multi-agent control system based on the fuzzy inference system for a group of two wheeled mobile robots executing a common task. Application of our approach lies in the control of robotic formations moving in the plane such as a group of guard robots taking care of one area and dealing with potential intruders. II. PROBLEM STATEMENT First task for the group of the robots is to maintain optimal formation in the specified area. This means to assign a position for each robot to take care of. We might define formation control problem as follows: Given is a group of n two wheeled mobile robots that are executing common task in the plane with obstacles. The task is to navigate each robot to the position optimal according to the task. In the area of surveillance, goal is means to cover all areas with expected intruder occurrence. If the map of the area is available and we have some other preliminary information about places of potential intrusion, resolving of the sub-optimal positions for the guard robots is not a very big problem. In this work we want to focus on the following scenario: When intruder appears, the task for the guard robots is to contact the intruder by coming near. This could be formalized as a pursuit-evasion problem. The task for the team of guard robots is then to catch the intruder in the lowest possible time. There can be given some constraints for the positions of the robot such as obstacles or expectation of second intruder. III. STATE OF THE ART According to the available literature, pursuit evasion problem is defined as follows: Pursuit-evasion task is to control a group of mobile robots to catch one or more evading agents in finite time. Pursuit-evasion evasion can be divided according to the number of agents: • One evader, one pursuer – in this case, the game is reduced to the simple navigation task • One evader, more pursuers – this case is the most covered by the literature, because in can be easily transformed to following cases • More evaders, one pursuer – single agent catches more evading entities • More evaders, more pursuers – the most general type of the game Interesting question is what does it mean to catch a evader. There are two known approaches: • Catching by approaching to the critical distance • Catching by “spotting” evader in the visibility region We recognize several types of pursuer visibility regions. Some approaches consider pursuer's visibility region as a thin laser line, and divides pursuers according to the number of lines available (k-searcher). Others resolve visibility regions as angular area, denoted by its size (δ-searcher). Most of the available papers solve this problem by formalizing pursuit evasion problems using these methods: Graph representation: The pursuit evasion game happens in the oriented graph. [3] Nodes represent available agent positions and arcs are connections between them. This is the most basic and historically first form. Evader is caught, when it occupies the same node as pursuer. Probabilistic representation. Environment is modeled using grid world. Every square has assigned probability of the evader occurrence. Evader is caught, when pursuer approaches to the critical distance [4, 5] Polygonal representation: Area is simplified to the polygonal objects and agents move on its arcs with the given speed. In this case, we consider k-searcher or δ- searcher. [1, 7, 11] 978-1-4244-2106-0/08/$25.00 © 2008 IEEE. 273