Reinforcement Learning-Based Human Kinetics Estimation from Kinematics

Human kinetics represents a pivotal research domain within rehabilitation biomechanics. Inverse dynamics serves as a prevalent method for calculating the joint torque of human movement, utilizing ground reaction force and kinematics. However, obtaining ground reaction force can be challenging for movements characterized by expansive activity spaces or frequent foot-ground interactions. In this study, we introduce a novel joint moment estimation scheme capable of generating biologically plausible torque profiles to track motion based on reference kinematics data. This scheme employs reinforcement learning in conjunction with a cascade proportional-derivative controller. To enhance the accuracy of joint torque and ground reaction force estimation, a data fusion scheme is devised, integrating biomechanical factors into the reward function. The efficacy of the proposed method is validated through simulations across various running speeds. An ablation study underscores the advantages of our controller design. The proposed algorithm is poised to serve as a complementary and alternative approach to inverse dynamics in human movement analysis.