Optimizing robot striking movement primitives with Iterative Learning Control

Highly dynamic tasks that require large accelerations and precise tracking usually rely on precise models and/or high gain feedback. While movement primitives allow for efficient representation of such tasks from demonstrations, the optimization of the required motor commands for systems with inaccurate dynamic models remains an open problem. To achieve accurate tracking for such tasks, we investigate two related Iterative Learning Control update laws and present a variant suited for optimizing hitting movement primitives. The resulting algorithm generalizes well to different initial conditions and naturally addresses striking movements where reaching specific velocities at certain positions is crucial. We evaluate the performance of our approach in a simulated putting task as well as in robotic table tennis, where we show how the striking performance of a seven degree of freedom anthropomorphic arm can be optimized. Our final implemented algorithm compares favorably with two state-of-the-art approaches.