关键词: 双足步行, 深度强化学习, 运动控制

Abstract: To solve the problems of motion instability of humanoid robots in variable uncertain, unstructured terrain and the low accuracy motion control, intelligent posture motion control algorithm is proposed. The deep reinforcement learning based continuous motion and continuous state space is applied to posture control, and the humanoid robot motion intelligent posture controller is established. Aiming at the problems of less sample and low efficiency of physical prototype training, the identification robot model is present to perform offline pre-training of the posture controller as a prior knowledge for continuous learning and in the real physical environment, improve the training efficiency in the later stage. The optimized robot posture controller is applied to the motion control of the robot. Compared with the robot motion with PID controller, MPC controller and PID+MPC controller, the standard deviation of the upper body pitch posture trajectory tracking error of the robot is reduced by 60.97%, 46.36%, 23.98% in the environmental transitional walking test, respectively. In the walking test of ground obstacles, the standard deviations of the trajectory tracking errors of the robot's upper body pitching posture are reduced by 60.38%, 26.38% and 9.52%, respectively.

Key words: bipedal walking, deep reinforcement learning, motion control