Optimal Control of Multi-phase Movements with Learned Dynamics

In this paper, we extend our work on movement optimisation for variable stiffness actuation (VSA) with multiple phases and switching dynamics to incorporate scenarios with incomplete, complex or hard to model robot dynamics. By incorporating a locally weighted nonparametric learning method to model the discrepancies in the system dynamics, we formulate an online adaptation scheme capable of systematically improving the multi-phase plans (stiffness modulation and torques) and switching instances while improving the dynamics model on the fly. This is demonstrated on a realistic model of a VSA brachiating system with excellent adaptation results.