In-hand manipulation via deep reinforcement learning for industrial robots

Abstract

Robotics manipulation is still a challenge in many scenarios, especially when the orientation of the tool or part is determinant for task success. A study with pivoting, an in-hand manipulation technique, was conducted, which consists in re-orienting the part around one rotational axis without dropping. It gets more complex in industrial robots, which are position controlled and often the user does not have access to the dynamic parameters. Deep reinforcement learning has been successful with model free approaches as it learns the behavior of the whole system. A simulated experiment for pivoting was conducted for part alignment to a desired angle with a one degree of freedom robot and a parallel gripper. Position control and torque control were simulated and compared.