Classifier-Aided Maximization of Feasible- Error-Region for Robust Manipulation Learning

Abstract

In this study, we propose an algorithm that enables the active learning of a manipulation skill that is robust to the estimation error of an object s position. First, we propose the concept of a feasible error region (FER) unique to each manipulation parameter. As the name implies, FER is a region of estimation errors even with which the manipulation still succeeds. We introduce the volume of FER as a robustness measure for manipulation parameters. Then, we develop an active learning algorithm capable of simultaneously realizing the volume maximization and estimating the FER corresponding to the optimal parameter. The obtained FER can be directly used to assess the success probability of the manipulation. As well as numerical experiments using toy models, we apply the algorithm to a robotic task in the physics simulator and two real world tasks. We demonstrate the validity of our approach, and discuss its limitations.