Tactile-based screw picking using reinforcement learning



Although robotic picking has been studied extensively, picking small objects such as screws is still challenging. The main reason for this is the difficulty in acquiring depth images used in conventional systems. In this study, we proposed a robotic system that can pick a single screw by relying only on the tactile sense of the fingertips. Without the use of images, it is difficult to know the status of screws until they are grabbed. However, by using flexible fingers, the robot is able to grasp multiple screws without any complicated control. Then, using reinforcement learning, the robot acquired the human-like behavior of dropping unnecessary screws by sliding its fingers. The results of this study are expected to serve as a basis for using robots to replace tasks that are still performed by humans in factories such as kitting.

Matthew Ishige, Takuya Umedachi, Yoshihisa Ijiri, Tadahiro Taniguchi, Yoshihiro Kawahara

Contact: mishige@akg.t.u-tokyo.ac.jp