A multi-agent reinforcement learning algorithm for collaborative skill transfer in physical education

Abstract In order to solve the problems of poor inter-limb coordination, delayed temporal adjustment, and the lack of stability in the transfer of skills, as well as the challenge of unifying multimodal motion representation and quantification of cross-limb contributions, this paper proposes a multi-agent reinforcement learning model tailored to physical education. Based on multi-agent Markov decision processes, this model integrates video stream, skeleton stream, and kinematic sequence coding. It incorporates spatio-temporal graph attention, message passing, counterfactual credit allocation, and confidence region stabilization control mechanisms to form a “representation—decision—feedback—closed-loop optimization” algorithmic chain. Experimental results show that the model’s pose error () is 2.33 pixels, outperforming the 3.87 pixels of single-agent, 3.42 pixels in a multi-agent system, and 3.05 pixels in a globally shared multi-agent system.