Physics-Constrained Reinforcement Learning for Individualized and Biomechanically Feasible Task-Motion Optimization

Abstract: The construction industry faces high musculoskeletal disorder rates due to repetitive, awkward task-postures. Current manual corrections and static optimizations often ignore individual anthropometry and whole-body biomechanical interdependencies, leading to physically unfeasible results. This study explores Reinforcement Learning (RL) for tailored posture optimization by developing a pipeline that overcomes technical barriers like reward instability and high computational costs. The proposed technique predicts joint-space residuals via a distilled, risk-aware RL policy within a physics simulator. Testing in construction-like scenarios showed significantly reduced mean risk and high-risk spikes while maintaining task fidelity. These findings demonstrate that RL can provide personalized, executable ergonomic interventions.