Application Of Reinforcement Learning In Path Planning Of Manipulator

Two-arm manipulator is a special robot. In process of path planning, manipulator's arms are closely linked. Manipulator is not simplified as a point or a circle. In this paper, a path planning method of manipulator based on reinforcement learning (RL) is proposed which is to be used in the two-dimension path planning problem of two-arm manipulator. The simulation results show that this method is quite efficient.RL gets optimal policy through trial-and-error and interaction with environment. As an unsupervised learning method, RL learns directly from the feedback of environment, which enable the system with RL to learn online and be adaptive to the varying environment. Its properties of self-improving and online learning make RL become one of the most important machine learning methods. Unlike that supervised learning needs input-output pairs RL only uses reward signals to improve its behavior.In this paper, the foundation and structure of RL is surveyed at first. Markov decision process (MDP) is introduced as the theoretical model of RL. The main algorithms of RL are introduced systematically. Markov games are introduced secondly as the theoretical model of multi-agent RL. The basic algorithm of multi-agent RL is improved. The new algorithm can cut down the redundant state information, so that the composition intensity of learning space is decreased and the convergence of the learning course is accelerated. Finally, a new structural credit assignment (SCA) method to assign reward signals to every agent by proportion is proposed, with which the multi-agent RL algorithm is used successfully in the path planning of manipulator.