The Art of Manipulation: Learning to Manipulate Blindly

Abstract

Performing skillfull manipulation is a very challenging task for robots. So far, even experts could barely program them to e.g. perform the well known peg-in-hole problem in the real world. Autonomously acquiring such skills, let alone generalizing them to new tasks, is still a major challenge. Typically, manipulation learning is approached with the help of large computation power, very long learning times, or possibly both. However, the performance achieved up to now is still far from human performance. We show the results of our new paradigm to robot manipulation. It bridges and unifies basic motor control, simple and complex manipulation strategies and high-level manipulation planning. The robots show autonomous skill learning, intra-class and inter-class generalization of insertion skills at human-level performance.