Tele Running - Energy Efficient Locomotion for Elastic Joint Robots by Imitation Learning

Abstract

This thesis presents an imitation learning approach to energy-efficient trajectory generation for elastic, legged robots. The trajectories are generated by teleoperation with force feedback. The presented framework allows an operator to achieve locomotion on an one-leg hopper by controlling its foot tip. The force feedback is designed to assist the operator to find gaits which exploit the natural harmonics of the hopper and thus improve energy efficiency. The resulting trajectory is approximated, parameterized, and replayed on the robot. The operator achieves a cost of transport of 0.25 at 0.63 m/s, considering the mechanical energy. Black-box optimization is used to keep this value with varying hardware parameters, such as different foot-tip stiffness. A reinforcement learning algorithm stabilizes lateral movement by active balance in simulation. Learning on hardware shows an improvement in stability. The concept is extended to multi-legged robots by teleoperating the two feet of the biped DLR C-Runner in simulation. The force feedback assists the operator to find stable gaits where the center of mass does not leave the support polygon of the feet. On both systems, the presented teleoperation framework utilizes the human's capability of estimating the properties of non-linear dynamics by designing appropriate haptic feedback.