Compliant assembly method for moving targets based on reinforcement learning

ZHANG Shuting; WAN Xiaojin

doi:10.16578/j.issn.1004.2539.2026.02.010

Abstract：[Objective]Aiming at the problem that most research on robotic autonomous assembly focuses on static targets while there is insufficient research on moving target assembly,a variable parameter control method for visual impedance controllers was proposed to improve the adaptability and compliance of robots in target tracking and moving assembly in dynamic environments.[Methods]Firstly,a 2nd-order impedance controller was constructed in the feature space by integrating force feedback and visual feedback information to provide a control foundation for dynamic assembly.Secondly,Monte Carlo dropout(MCD)was used as the probabilistic dynamic model to improve the probabilistic inference for learning control optimization(PILCO)algorithm,balancing the ability of state uncertainty reasoning and computational efficiency.Finally,the impedance parameters of the controller were adaptively adjusted using feature errors and robot joint position information as the observation space to optimize dynamic tracking and assembly performance.[Results]The results of simulations and test bench verifications indicate that compared with the original Gaussian process model,the MCD model retains the ability of state uncertainty reasoning while significantly reducing the training time(the training time for 50 rounds is reduced from 43.50 h to 12.82 h);the assembly success rate is increased from 94.5%to 98.0%,and the average assembly time is shortened from 5.820 s to 3.253 s.The overshoot is greatly reduced and the tracking response is more timely,providing a reference for the compliant assembly of moving targets.