With the further research of robotics system, the focus of robotics research has shifted from the signle function and not independent ablity robots to the mobile robots that have a certain mobility and intelligence in natural environment. Intelligent robot has attracted a large number of researchers'attention in the word, and multi-robot system arising more and more attention becomes a major direction of robotics research. The method of multi-robot path planning is one of the most important issues of multi-robot research.Multi-robot path planning using multi-robot system as an object, finds an optimal path for each robot in the same workspace, and ensures that there is no collision between each robot. As the multi-robot path planning is a premise to perform tasks efficiently and reasonable, it's essential to improve the efficiency of the robot. How to choose a reasonable and a best path is a very important research topic.Firstly, this paper introduces the multi-robot system, the basic theory of multi-robot system, key technologies and applications etc. The following is introducing the analysis of path planning algorithms and their advantages and disadvantages. This paper also describes the reinforcement learning, hierarchical reinforcement learning, fuzzy logic and basic theory of SOM network. With the increase of robots'number and the complexity of the environment, the state space grows exponentially, but how to achieve a continuous state space, best approximation, reinforcement learning agents must have the Pan-ability. Because of neural network's ability of parallel computing, fault tolerance and non-linear approximation, the method of using controller combined by the SOM network and fuzzy logic to classify the reinforcement signals is proposed, achieving the multi-robot path planning. The system overcomes the fuzzy control disadvantages of depending entirely on the parameter adjusting and not having a memory function, improves the problems of relatively slow process of strategy searching in reinforcement learning, increases the robot's self-learning ability and meets the real-time requirement of system. The result of simulation proves the study is effectiveness and feasibility. |