Active learning for robot manipulation

Abstract

This paper describes a novel application of active learning techniques in the field of robotic grasping. A vision-based grasping system has been implemented on a humanoid robot. It is able to compute a set of feasible grasps and to execute any of them and measure their actual reliability. An algorithm aimed at predicting the performance of an untested grasp using the results observed on previous similar attempts is presented. The previous experience is stored using a set of vision-based grasp descriptors. Moreover, a second algorithm that actively selects the next grasp to be executed in order to improve the predictive quality of the accumulated experience is introduced. An exhaustive database of experimental data is collected and used to test and validate both algorithms. 1