Compliant Locomotion Control for a Quadruped Robot with Damper Coefficients Assigned by Reinforcement Learning

摘要

Since quadruped robots have been considered to be effective on rough terrain, they have been studied in various fields such as rescue and planetary exploration. In the legged locomotion literature, researchers mostly focus on mechatronics hardware and actuation systems of quadruped robots. Nevertheless, such systems may be questionable from the stability point of view, in particular, for dynamic walking and running. Therefore, to develop high locomotion capabilities to adjust to different type of terrains, we propose a new type of compliance control based on reinforcement learning. In this method, compliance control parameters of each joint can be adjusted by external forces acting on the robot feet. Due to this adaptive mechanism, dynamically balanced jumping motion and trotting quadruped locomotion on the rough terrain can be realized. In order to demonstrate the efficiency of the proposed method, jumping and trotting experiments were conducted on our quadruped robot. As a result, we obtained periodic, continuous and repetitive jumping and trotting cycles in which dynamic balance was ensured.