Learning Similar Dynamic Motions Incrementally Using Joint Trajectories Obtained by Quasi-Static Execution

Abstract

We propose a new supervised learning and synthesis framework for fast, complex motor tasks for redundant robots. A statics-based task-space controller acts not only as a full-body motion control module, but also as a module to generate synergistic joint motion patterns for redundant systems. Similar, but faster motions are incrementally synthesized by superposing the task-space controller output and stiffness around the joint trajectories with the modified speed, while iteratively learning the dynamics and joint stiffness according to the L2 norm of the task-space error. We demonstrate the proposed framework by simulating a balanced fast squat on a simple humanoid model.