A fast transfer reinforcement learning model for transferring force-based human speed adjustment skills to robots for collaborative assembly posture alignment

•A human-robot collaborative control model with human skills is proposed.•Learning human speed adjustment skills for responding to external force changes.•The control model intelligently converts human EMG signals into robot adjustment.•The Data-driven collaborative control model avoids control parameter selection. Human-robot collaboration demonstrates significant autonomy and flexibility, making it highly suitable for personalized and adaptable production tasks. However, the disparity between human-expected collaboration speed and actual robot collaboration speed diminishes the accuracy and comfort of human-robot interactions. This paper addresses this challenge by introducing a collaborative control method that leverages a transfer reinforcement learning algorithm to acquire human instinctive speed adjustment skills based on Electromyography (EMG) signals and human joint angle data, aimed at enhancing robot collaboration capabilities. Specifically, to address the challenge of time misalignment and feature mismatch between EMG signals and joint angle data in recognizing human posture adjustment intentions, with the help of a temporal dilation convolutional feature fusion network, a human posture intention inference model is proposed. Additionally, to achieve a balance between precise tracking and comfortable collaboration in human-robot interaction, and to mitigate abrupt changes in interaction forces, a data-driven collaborative control strategy is proposed. This control strategy intelligently converts human EMG signals into robot adjustment commands. To reduce training costs and model complexity, a transfer reinforcement learning model with multi-objective optimization capability is proposed to achieve transfer and generalization of human speed adjustment skills across multi-dimensional human-robot collaborative operation tasks. Finally, taking the posture alignment of multi-peg-in-hole assembly as an example, the proposed collaborative control method is experimentally verified. Experimental results show that compared to the mainstream adaptive impedance control method, the proposed collaborative control method with human speed adjustment skills reduces the equivalent output torque by 16.7 %, assembly time by 7.8 %, and assembly error by 33.0 %, effectively enhancing collaboration performance.