Biomechanics-driven deep reinforcement learning improves Wushu skill acquisition in physical-education students over an eight-week controlled intervention

Abstract Wushu instruction still rests on the coach’s eye and on repeated demonstration, an approach that resists quantification, delivers feedback only after a routine ends, and cannot retune difficulty for each learner. We therefore asked a deliberately narrow question: does a training paradigm in which biomechanically grounded features drive a reinforcement learning policy improve skill acquisition more than conventional coaching? Four directional hypotheses, stated before analysis, framed the test. Open-source inertial sensing was fused with markerless pose estimation to recover joint kinematics and inverse-dynamic torques, and a joint-angle-torque coupling scheme encoded those descriptors as a structured state tensor. A proximal-policy-optimization agent, driven by a composite reward and a dynamic difficulty-adjusting function, selected the training difficulty and the feedback shown to each learner. Sixty physical-education students completed an eight-week controlled intervention that compared the adaptive system with conventional coaching, with assessments at weeks 0, 4 and 8. Across sessions the instrumented measures proved reliable, with intraclass correlation coefficients from 0.82 to 0.91, and expert scoring showed good-to-excellent agreement both between raters, 0.91, and within raters, 0.93. Linear mixed-effects models with a random intercept for participant returned significant group-by-time interactions on all four skill dimensions: movement accuracy, $$F$$ (2, 116) = 18.42, $$p$$ < 0.001, partial $${\eta }^{2}$$ = 0.24; movement fluency, $$F$$ (2, 116) = 14.68, $$p$$ < 0.001, partial $${\eta }^{2}$$ = 0.20; force-control rationality, $$F$$ (2, 116) = 21.17, $$p$$ < 0.001, partial $${\eta }^{2}$$ = 0.27; and rhythmic stability, $$F$$ (2, 116) = 11.31, $$p$$ < 0.001, partial $${\eta }^{2}$$ = 0.16. The first three sit clearly above the conventional large-effect benchmark, whereas rhythmic stability only just clears it, and we read it accordingly. At the post-test the adaptive group exceeded the control group by 10.50 points on accuracy (95% CI 4.14 to 16.86, $$d$$ = 0.85) and by 11.60 points on force control (95% CI 5.06 to 18.14, $$d$$ = 0.92). Joint-angle variability also settled at a reduced plateau in the adaptive arm, and every within-participant change in an instrumented parameter exceeded the minimal detectable change of its own instrument. The convergence and transfer results point to the quality of the state encoding as a decisive design factor, although network depth itself was never varied, and the prediction that lower-baseline learners would gain most was not borne out. Because feedback modality and difficulty control were manipulated together, their separate contributions cannot be isolated in this design; the findings therefore support the packaged paradigm rather than either component on its own, and transfer to other coordination-intensive sports or to rehabilitation remains an untested conjecture. Trial registration : The trial was registered retrospectively rather than prospectively. Registration with a public clinical trial registry was applied for once data collection had ended, and the registration identifier will be supplied to the editorial office as soon as the registry record is issued; the absence of prospective registration is stated as a limitation in Section “Discussion”.