Reinforcement Learning-Based Impedance Learning for Robot Admittance Control in Industrial Assembly

摘要

Industrial assembly is essential for manufacturing, and robots have been broadly applied to assembly tasks. The conventional impedance control with fixed impedance parameters may restrict the compliance and interactivity of robots, posing a threat to the safety of robots and environments during assembly tasks. This paper presents an impedance learning method using reinforcement learning (RL) for robot admittance control in industrial assembly. A model-free RL method termed twin delayed deep deterministic policy gradient is utilized to tune stiffness parameters, guaranteeing the safety of robots and environments during the completion of assembly tasks. The proposed method is applied to a collaborative robot to complete the peg-in-hole task, a benchmark problem in industrial assembly. Simulation results show that the proposed method achieves a trade-off between force response and tracking accuracy, and performs better than the fixed admittance control in terms of safety and efficiency.