Motion Control for Biped Robot via DDPG-based Deep Reinforcement Learning

In the study of the passive biped robot, the avoidance of fall over is always an important direction of the research. In this paper, we propose Deep Deterministic Policy Gradient (DDPG) to control the biped robot walk steadily on the slope. For improve the speed of DDPG training, the DDPG used in the paper is improved by parallel actors and Prioritized Experience Replay (PER). In the simulation, we control different initial states that cause the biped robot to fall over. After the control, the biped robot can walk stably, which indicating that DDPG can effectively control the fall over of the biped robot.