강화학습 기반 OpenSim 자동 스케일링을 통한인체 운동 분석 정확도 향상 기법

Abstract

Precise musculoskeletal scaling is vital for reliable human motion analysis, yet traditional scaling methods are often performed manually, which is time-consuming and expert-dependent. To address these limitations, this study introduces an automatic scaling framework that combines the OpenSim gait2354 model with a Twin Delayed Deep Deterministic Policy Gradient (TD3) reinforcement learning agent. Using 100 static OpenCap trials, subject-specific RL environments were designed and trained with a reward minimizing total errors, RMS, and peak errors. Following over 2,000 episodes, errors significantly reduced: mean total errors from 0.00473 to 0.00168 m² (-64.5%), RMS from 14.7 to 8.8 mm (-40.1%), and peak errors from 27.6 to 15.9 mm (-42.4%). This satisfies OpenSim scaling guidelines (<10 mm RMS, <20 mm peak). The framework can be expanded to diverse anthropometries, effectively automating the scaling process and showing potential for online fine-tuning with new subjects.