Motions obtaining of multi-degree-freedom underwater robot by using reinforcement learning algorithms

摘要

This paper deals with motions obtaining of an underwater robot arm which have multi-degree of freedom by using reinforcement learning algorithms. A natural gradient Actor-Critic algorithm which uses Eligibility Traces is applied to the robot arm. In this algorithm, motion planning problems are modeled as finite state Markov decision processes. The robot arm is developed to have 4 joints, each joint consists 1 servo motor. The experiment results show the robot arm successfully learning to swim by feasible learning steps.