Biophysical Models With Adaptive Online Learning for Direct Neural Control of Prostheses

Abstract

Direct neural control of multi-articulating prosthetic hands is critical for achieving dexterous manipulation in unstructured environments. However, such control - predicting continuous movements over independent degrees of freedom - remains confined to research settings. In contrast, pattern recognition systems are widely employed for their simple, user-friendly training procedures, though their limitation to a set of discrete whole-hand poses restricts functionality. To bridge this gap, we designed a a direct neural controller and a training procedure designed to support adaptive retraining, enabling users to improve controller predictions or incorporate new movements using a single RGB camera. It explicitly models musculoskeletal dynamics and employs a neural network-based method for motor intent disambiguation, which we term "synergy inversion". The defined dynamics constrain the predicted kinetics and kinematics to a physiologically realizable manifold, while synergy inversion can capture nonlinear patterns of muscle coactivation missing from traditional musculoskeletal models. In experiments with eight biologically intact participants and two individuals with unilateral transradial amputation, the proposed paradigm predicted trajectories for seven degrees of freedom and improved performance through online learning, achieving lower error than both purely neural and purely biophysical baseline models. This work represents a step toward the adoption of direct neural control of upper extremity prostheses in real-world settings, offering the flexibility of pattern recognition training within a more performant control framework.