Safe peg-in-hole automatic assembly using virtual guiding force: A deep reinforcement learning solution

This paper proposes an automatic peg-in-hole assembly method based on deep reinforcement learning, incorporating a feedforward virtual guiding force with safety considerations. Unlike traditional approaches that involve positional trajectories, our method draws inspiration from human dragging movements and utilizes feedforward virtual guiding forces as actions. This enables the reinforcement learning agent to drag the end effector to achieve peg-in-hole assembly and effectively addresses the safety issues caused by random actions throughout the training and testing processes. Experimental validation involves testing scenarios with pegs and holes featuring varying chamfers and clearances, as well as different levels of positioning uncertainty and initial search positions. The experiments demonstrate that our approach not only tackles the safety challenges but also exhibits good performance in cylindrical peg-in-hole tasks with initial positioning uncertainty and different chamfer/clearance structures, achieving high success rates.