What is it about?

This work studies how muscle signals from the residual limb can be used to predict the motion of a virtual hand. We analyzed recordings from five participants with below-elbow amputations who took part in VR-based pre-prosthetic training. During the training, six surface EMG sensors recorded muscle activity from the residual limb, while the target was to predict 20 continuous finger joint angles. We compared two ways to represent the EMG signal before prediction. One used a CNN feature extractor. The other used covariance features, which describe how the EMG sensors change together within a short time window. The covariance-based approach predicted virtual finger motion more accurately than the CNN baseline, while using fewer trainable parameters, less memory, and sub-millisecond processing time per window.

Featured Image

Why is it important?

Real-time VR training systems need movement prediction that is accurate, fast, and lightweight. This is hard because the system has to predict many finger-joint angles from only a few noisy muscle sensors placed on the residual limb. Our results show that a compact covariance-based representation can work better than a larger learned feature extractor for this task. That matters for myoelectric interfaces, where models often need to run under strict latency, memory, and hardware constraints.

Read the Original

This page is a summary of: Poster: Geometry-Aware Real-Time Finger Decoding from Residual-Limb sEMG, June 2026, ACM (Association for Computing Machinery),
DOI: 10.1145/3812835.3814867.
You can read the full text:

Read

Contributors

The following have contributed to this page