Learning Musculoskeletal Dynamics with Non-Parametric Models

Gait control of bipeds has been a key challenge in robot locomotion for decades. A key component toward controller design is a good model of the robot and its interactions with the environment. For systems with highly compliant and complex actuation, however, deriving the models for the essential system components is quite tedious. In this paper, we propose to learn non-parametric models for a human inspired musculoskeletal biped, whose nonlinear actuation dynamics and passive tendon elasticities make modeling very challenging. We present first promising results that hint at the usefulness of datadriven modeling approaches in the context of legged robots.