Policies for Goal Directed Multi-Finger Manipulation

We present a method for one-handed task based manipulation of objects. Our approach uses a mid-level multiphase approach to break the problem into three parts, providing an appropriate control strategy for each phase and resulting in cyclic finger motions that accomplish the task. All motion is physically based, and guided by a policy computed for a particular task. The exact trajectory is never specified as the goal of our different tasks are concerned with the final orientation and position of the object. The offline simulations used to learn the policy are effective solutions for the task, but an important aspect of our work is that the policy is general enough to be used online in real time. We present two manipulation tasks and discuss their performance along with limitations.