Learning Task-Specific Dynamics to Improve Whole-Body Control

In quadratic program based inverse dynamics control of underactuated, free-floating robots, the desired Cartesian reference motion is typically computed from a planned Cartesian reference motion using a feed-forward term, originating from a model, and a PD controller. The PD controller is there to account for the discrepancy between the real robot and the model. In this paper we show how we can reduce the required contribution of the feedback controller by incorporating learning of Cartesian space reference accelerations. Thus, we i) improve the execution of the given specific task, and ii) offer the means to reduce feedback gains, providing for greater compliance of the system. With a systematic approach we also reduce heuristic tuning of the model parameters and feedback gains, often present in real-world experiments. In contrast to learning of task-specific joint-torques, which might produce a similar effect, our approach is applied in the task space of the humanoid robot. Simulated and real-world results on the lower part of the Sarcos Hermes humanoid robot demonstrate the applicability of the approach.