Affiliations:
School of Power and Mechanical Engineering, Wuhan University, Wuhan, P.R. China
Correspondence:
[*]
Corresponding authors. Xiaohui Xiao and Zhao Guo, School of Power and Mechanical Engineering, Wuhan University, Wuhan, 430072, P.R. China. Tel.: +86 27 6877 2247 (X. Xiao); Fax:+86 27 6877 2040 (X. Xiao); E-mails: xhxiao@whu.edu.cn (X. Xiao) and guozhao@whu.edu.cn (Z. Guo).
Abstract: Passive dynamic walking (PDW) has attracted much research attention due to its humanoid and energy efficient gaits. However, walking control of the passivity-based biped robot inspired by PDW still remains a challenge, for PDW is sensitive to disturbances. An walking controller is essential for practical passivity-based biped robots in real environments. This paper presents a deep reinforcement learning (DRL) controller based on deep Q network for planar passivity-based biped robot, to learn policies directly from inputs for bipedal walking task. First, the intelligent controller using deep Q network is trained, with PDW as reference trajectory. The learning experience from PDW could be helpful to implement a natural looking and energy-efficient gait. Then the trained deep Q network is utilized as the walking controller. Simulation results show that the DRL controller based on deep Q network makes the planar biped robot walk against original value disturbance, on different slope, level ground and varying slopes. The controller this paper presented could be used to improve the versatility of the passivity-based biped robot.
Keywords: Biped robot, deep Q network, deep reinforcement learning, passive dynamic walking, Q learning
