Active Tactile Object Recognition by Monte Carlo Tree Search.

This paper considers the problem of object recognition using only tactile information. The focus is on designing a sequence of robot hand grasps that achieves accurate recognition after few enclosure contacts. It seeks to maximize the recognition probability and minimize the number of touches required. The actions are formulated as hand poses relative to each other, making the algorithm independent of small object movements and absolute workspace coordinates. The optimal sequence of actions is approximated by Monte Carlo tree search. We demonstrate active tactile recognition results in physics simulation and on a real robot. In simulation, most object instances were recognized within 16 moves. On a real robot, our method correctly recognized objects in 2--9 grasps and outperformed a greedy baseline.