Reinforcement learning-based controller for quadruped locomotion over compliant terrain

Abstract

This thesis presents the implementation and evaluation of a reinforcement learning-based controller for the Unitree Go1 EDU over compliant terrain. A complete kinematic analysis of the Go1 is derived to detect the forces acting on the feet of the robot during locomotion in varying environments. This proved to be successful in the identification of a rigid versus compliant surface. A soft foam pad serves as the compliant surface throughout this research. To achieve successful locomotion over this surface, we have proposed and deployed a learning-based controller to the Unitree Go1 that is trained in a simulated compliant environment. The proposed controller is analyzed in comparison to the baseline reinforcement learning-based controller trained in a general environment. Both controllers are evaluated over the soft foam surface with a motion capture system tracking passive markers on the robot's base. Metrics, such as position, velocity, acceleration, and jerk of the center of mass (COM) of the robot, are extracted to provide a thorough comparison between the two controllers. Experimental results display very similar performance metrics between both controllers. There is however a slight performance decrease in the proposed controller in two out of the three accounts when comparing acceleration and jerk over compliant terrain. The scope of this work serves to provide insightful information into both the control methods and gait strategies of quadrupeds in compliant environments.