Generation of Whole-body Locomotion of a Humanoid Robot by Reinforcement Learning

Abstract

This paper presents a method for locomotion generation of a humanoid robot walk by reinforcement learning (RL). Specifically, the RL is designed to learn a policy for outputting the joint torques that perform periodic combined motions of limbs, aiming at acquiring a whole-body locomotion. The effectiveness of the learning method is validated on the MuJoCo simulator using the small humanoid robot ROBOTIS-OP3. We report the results obtained by adjusting the reward terms accounting for performance for biped and quadruped locomotions and other parameters.