Learning Efficient Omni-Directional Capture Stepping for Humanoid Robots from Human Motion and Simulation Data

Abstract

Two key questions in the context of stepping for push recovery are where to step and how to step there. In this paper we present a fast and computationally light-weight approach for capture stepping of full-sized humanoid robots. To this end, we developed an efficient parametric step motion generator based on dynamic movement primitives (DMPs) learnt from human demonstrations. Simulation-based reinforcement learning (RL) is used to find a mapping from estimated push parameters (push direction and intensity) to step parameters (step location and step execution time) that are fed to the motion generator. Successful omni-directional capture stepping for 89 % of the test cases with pushes from various directions and intensities is achieved with minimal computational effort after 500 training iterations. We evaluate our method in a dynamic simulation of the ARMAR-4 humanoid robot.