Learning-Based Inverse Dynamics of Human Motion

Abstract

In this work we propose a learning-based algorithm for the inverse dynamics problem of human motion. Our method uses Random Forest regression to predict joint torques and ground reaction forces from motion patterns. For this purpose we extend temporally incomplete force plate data via a direct Random Forest regression from motion parameters to force vectors. Based on the resulting completed data we estimate underlying joint torques using a modified physics-based predictive dynamics approach. The optimization results for model states and controls act as predictors and responses for the final Random Forest regression from motion to joint torques and ground reaction forces. The evaluation of our method includes a comparison to state-of-the-art results and to measured force plate data and a demonstration of the robust performance under influence of noisy and occluded input.