Walking Motion Generation of Bipedal Robot Based on Planar Covariation Using Deep Reinforcement Learning

Recently, a deep reinforcement learning approach has been proposed as one of the control methods for legged robots. In this paper, we describe the walking motion of a bipedal robot using deep reinforcement learning and define a human-like biped gait as a gait in which the robot lands on its heels and takes off its toes. In this study, we focus on the kinematic synergy (planar covariation) pointed out by Borghese et al. as a characteristic gait of humans. Planar covariation is angular data plotted in the sagittal plane with the elevation angles at the thigh, shank, and foot on the three axes falling on a single plane. We propose this feature as the reward for reinforcement learning. By introducing this reward, the bipedal robot realized a human-like gait. We also consider that the proposed reward is one factor that characterizes human walking motion by comparing the learning results when this feature is not used.