Joint Torque Estimation for Index Finger using sEMG and Joint Deflection: Ensuring End-to-End Transparency

Abstract

Motion Copying Systems (MCS) have long been explored to reproduce human motion, yet the fundamental question of motion abstraction, i.e., simultaneous representation of both force and position for scalable skill transfer, remains under-addressed. Current methods, such as optical tracking methods, electromechanical devices, and wearable sensors, have limited generalizability and portability and lack a complete representation of motion dynamics. In addition, much research has focused on using surface electromyography (sEMG) to infer force, but it is generally modeled independently of motion, with most methods relying on black-box deep learning models. This paper proposes a novel motion abstraction framework that leverages sensor fusion of sEMG and flex sensors with an explainable artificial intelligence (xAI) approach to transparently estimate joint torque with motion compensation. A two-model formulation is created: Model A estimates the sEMG component associated with muscle activity for free motion, and Model B estimates the torque delivered to the environment. The estimated joint torque can then be used to deduce the external force applied by the finger during motion. Both these models are developed using the Element Description Method, which uses a genetic algorithm for parameter optimization. Sensor fusion of a flex sensor and sEMG makes it possible to generate a small database of motor skills that can be replayed on different robots and can be used to train human learners. Experimental results demonstrate that the proposed method achieves accurate estimation of the force that muscles apply to the environment while maintaining interpretability and transparency.