Compliant Robotic Assembly based on Deep Reinforcement Learning

Assembly with industrial robots in an unstructured environment is more and more desirable and going to play an indispensable role in modern manufacturing industries. The classical peg-in-hole robotic assembly has been extensively researched as a common industrial task, but high-precision robotic assembly, which means the assembly precision of mechanical parts required may exceed that of robots, remains an open problem. Meanwhile, the conventional impedance control method used in the current manufacturing, requires numerous parameters to be artificially tuned before deployment, tedious and awfully time-consuming. Moreover, it has a poor success rate. In this paper, we propose a novel method with off-policy, model-free deep reinforcement-learning for position-controlled robots to perform assembly tasks compliantly through training, and there is no need for manual tuning. Through conducting a series of comparative experiments, this method is proved to have a preferable performance and assembly efficiency.