Task-specific dynamics for robotic hand control

Humans employ context-specific internal models for their bodies and the world around them. They learn them through embodiment in the environment, by gathering data in a task-specific way. Similarly, robots operating under humanlike task constraints must have the ability to accommodate situations unknown to their designers. We demonstrate a method for the Anatomically Correct Testbed (ACT) robotic hand to use task-relevant data to build a reduced-dimensionality controller tailored to that task. The robotic hand encounters a novel task and models the combined dynamics of robot and environment. This is achieved without additional model complexity, and without prior knowledge of the task. We show the utility of this approach for playing a piano key, attaining a single-key trill at the maximum speed allowed by the piano dynamics. This task is chosen because it is relatively simple from a kinematic perspective, mostly involving flexion-extension, but is quite complex from a dynamic perspective, including a contact transition.