Abstract
Lower limb musculoskeletal dynamics simulation has been widely used to estimate the lower limb mechanics, but challenges such as heavy reliance on force plates, poor model generalization, and high computational load hindered its application in real-time robot control systems requiring rapid feedback and inference. This study proposed the Marker-GMformer model, a marker trajectories-driven deep learning model designed for efficient and accurate continuous prediction of lower limb kinematics and dynamics. By integrating prior knowledge with global-local and spatial-temporal features from the inputted marker coordinate time series, Marker-GMformer maintained high performance while reducing computational complexity. The model also demonstrated strong generalization, accurately predicting multi-joint kinematics, moments, and ground reaction forces (GRFs) across 13 different motion patterns. The predicted results were compared to those from musculoskeletal multibody dynamics simulations and force plates. Excellent performance was achieved with average Pearson correlation coefficients ( ρ ≥ 0.97 ) and low root mean square errors (RMSE = 1.95° for angles, RMSE = 0.036 body weight for GRFs, and RMSE = 0.099 N·m/kg for moments) across all patterns. The findings underscored the substantial promise of the proposed method for enabling real-time monitoring of human lower limb mechanics and delivering timely feedback to optimize the control of assistive robots.