Bio-inspired motion control of the musculoskeletal BioBiped1 robot based on a learned inverse dynamics model

Abstract

Based on the central hypothesis that a humanoid robot with human-like walking and running performance re- quires a bio-inspired embodiment of the musculoskeletal functions of the human leg as well as of its control structure, a bio-inspired approach for joint position control of the BioBipedl robot is presented in this paper. This approach combines feed- forward and feedback control running at 1 kHz and 40 Hz, respectively. The feed-forward control is based on an inverse dynamics model which is learned using Gaussian process regression to account for the robot's body dynamics and external influences. For evaluation the learned model is used to control the robot purely feed-forward as well as in combination with a slow feedback controller. Both approaches are compared to a basic feedback PD-controller with respect to their tracking ability in experiments. It is shown, that the combined approach yields good results and outperforms the basic feedback controller when applied to the same set-point trajectories for the leg joints.