Background
Direct neural control predicts continuous movement across the independent degrees of
freedom of a multi-articulating prosthetic hand, enabling far more dexterous manipulation
than the discrete whole-hand poses offered by conventional pattern recognition systems.
However, direct control has remained largely confined to research settings, while pattern
recognition remains popular for its simple, user-friendly training procedure. This project
set out to combine the flexibility of pattern-recognition-style training with the
performance of direct neural control.
Approach
We designed a direct neural controller, driven by surface EMG, that explicitly models
musculoskeletal dynamics: a biophysical muscle model and joint model constrain the predicted
kinetics and kinematics to a physiologically realizable manifold with a neural network to predict
virtual muscle activation from surface EMG. The full pipeline supports adaptive online retraining,
letting users correct for temporal drift or add new movements from a single RGB camera rather than
a full motion-capture setup.
In experiments with eight biologically intact participants and two individuals with
unilateral transradial amputation, the model predicted trajectories for seven degrees of
freedom and after perturbation returned to baseline performance within minutes of online learning.
References
[1] J. Gentinetta, M. F. Fernandez, J. Qiao, M. R. Gonzalez and H. M. Herr, "Biophysical
Models With Adaptive Online Learning for Direct Neural Control of Prostheses," IEEE
Transactions on Neural Systems and Rehabilitation Engineering, vol. 33, pp. 3201-3211, 2025,
doi:
10.1109/TNSRE.2025.3599114.