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

Area of Science:

• Biomedical Engineering
• Neuroscience
• Robotics

Background:

• Direct neural control of prosthetic hands is crucial for dexterous manipulation but remains largely in research settings.
• Current pattern recognition systems offer user-friendly training but are limited to discrete hand poses, restricting functionality.

Purpose of the Study:

• To develop a direct neural controller for multi-articulating prosthetic hands that bridges the gap between research-level control and practical application.
• To enable adaptive retraining for improved controller predictions and incorporation of new movements using a single RGB camera.

Main Methods:

• Designed a direct neural controller modeling musculoskeletal dynamics and employing "synergy inversion" for motor intent disambiguation.
• Utilized a neural network-based method to capture nonlinear muscle coactivation patterns.
• Implemented adaptive retraining with a single RGB camera for user-driven improvements.

Main Results:

• The proposed paradigm successfully predicted trajectories for seven degrees of freedom in both intact participants and amputees.
• Online learning improved controller performance, outperforming purely neural and biophysical baseline models.
• The system demonstrated lower prediction error compared to baseline models.

Conclusions:

• The developed direct neural controller offers a more performant control framework with the flexibility of pattern recognition training.
• This work advances the adoption of direct neural control for upper extremity prostheses in real-world applications.
• The "synergy inversion" method effectively models complex motor intent for prosthetic control.