M-A3C: A Mean-Asynchronous Advantage Actor-Critic Reinforcement Learning Method for Real-Time Gait Planning of Biped Robot

Bipedal walking is a challenging task for humanoid robots. In this study, we develop lightweight reinforcement learning method real-time gait planning of the biped robot. We regard bipedal as process in which robot constantly interacts with environment, judges quality control action through state, and then adjusts strategy. A mean-asynchronous advantage actor-critic (M-A3C) algorithm proposed to obtain continuous state space space, directly final without introducing reference gait. use multiple sub-agents M-A3C train virtual robots independently at same time physical simulation platform. Then transfer trained model actual reduce number training on robot, improve speed, ensure acquisition Finally, designed fabricated verify effectiveness method. Various experiments show that can achieve robot’s stable planning.