Contact Invariant Model Learning for Legged Robot Locomotion

Abstract

In this work we present a new formulation for learning the dynamics of legged robots performing locomotion tasks. Using sensor data we learn error terms at the level of rigid body dynamics and actuation dynamics. The learning framework deals with the hybrid nature of legged systems given by different contact configurations: We use the projection of the rigid body dynamics into a subspace consistent with the contact constraints. The equations of motion in such subspace do not depend on the contact forces, allowing to formulate a learning problem where force sensor data is not required. Additionally, we propose to use the columns of end-effector Jacobians as basis vectors, obtaining a model that generalizes across contact configurations. Both Locally Weighted Projection Regression and Sparse Gaussian Process Regression are used as supervised learning techniques. As application of the learned model, an inverse dynamics control method is extended. Hardware experiments with a quadruped robot show reduced RMS tracking error and a significant reduction in RMS feedback effort during base-only, walking, and trotting motions.