Walking parameters design of biped robots based on reinforcement learning

Abstract

Biped walking pattern is one of the most difficult problems in the humanoid robot area, and there exists no ideal algorithm for a generalized walking scenario. Recently, some methods have been proposed, including trajectory planning, passive dynamic walk. In this paper, based on the solution of inverse kinematics of a leg by combining analysis method with numerical method, trajectory planning method is used to implement the humanoid robot walking skill in a 3D simulation environment. In order to get the walking parameters automately, reinforcement learning is studied and implemented by the train system of Apollo3D program, and the training algorithm is well tested in the RoboCup3D simulation platform.