Manipulation Skill Acquisition for Robotic Assembly using Deep Reinforcement Learning

Abstract

Nowadays mobile manipulators are commonly used in assembly tasks, which can reach greater workspace but also cause more uncertainties. Uncertainty is one of the main factors affecting the quality and efficiency of assembly tasks. The framework of skill acquisition based on machine learning can make a good job of solving the uncertainties in mobile manipulation. In this paper, a method based on deep reinforcement learning is proposed to learn assembly skills. The Deep Deterministic Policy Gradient algorithm based on off-policy can scale to complex manipulation tasks. The safety constraint was set to the torque of each joint. The positive and negative reward function is designed to improve learning efficiency, which could be used in different assembly tasks. The deep neural network policies could be trained out efficiently on a real robot platform. This approach was performed on the KUKA iiwa robot arm equipped with 7-dof force/torque sensors. The results show that the robot could learn assembly skills properly with no prior knowledge.