Reinforcement-Learning-Driven Automatic Scaling in OpenSim for Enhanced Human Motion Analysis

Musculoskeletal simulation depends on accurate subject-specific model scaling, but the scaling step still introduces a major real-to-sim gap in human motion analysis. Manual scaling is time-consuming, subjective, and difficult to reproduce, while optimization-based auto-scaling can be sensitive to initial values and local minima. These limitations can propagate into downstream analyses such as inverse kinematics, inverse dynamics, and joint-torque estimation.

This project formulates OpenSim model scaling as a reinforcement-learning problem. Instead of directly relying on manual tuning or a single optimization run, an agent learns to search for 3D scale factors that minimize marker error between experimental and virtual markers. The framework adopts a TD3-based actor-critic architecture with a Transformer network so that subject-specific context such as height, weight, and sex can be incorporated into the scaling policy.

The Transformer-based RL model achieved faster convergence than a feed-forward architecture, reduced marker errors compared with AddBiomechanics, and accelerated elderly-cohort fine-tuning through OpenCap pre-training.