Active Learning of robust manipulation by feasible region maximization

Abstract

The objective of this paper is the proposal of an active learning algorithm for robust manipulation. To this end, we introduce the concept of the feasible region and propose the size of the feasible region as a robustness measure of a trajectory. Roughly speaking, the feasible region is a region of acceptable perception error when a certain trajectory is executed. We then propose a method to obtain a robust trajectory by maximization of the robustness measure. Maximization and estimation of the feasible region size are performed simultaneously by the active learning of a real robot. The great thing of our approach is that a feasible region reflects physical interaction between the gripper and the object while the corresponding trajectory is executed. We applied our method to the door opening task of microwave-oven. After active learning, the robot acquired a robust manipulation skill which takes advantage of physical interaction with the microwave-oven.