A generic non-invasive neuromotor interface for human-computer interaction

The 60-second view

What problem?

High-bandwidth neural input had needed invasive interfaces and decoders built for one person.

What did they do?

Collected sEMG from thousands of participants and trained generic models for gestures, navigation and handwriting.

What changed?

Demonstrated out-of-the-box generalisation across users, improved further by personalisation.

Why does it matter?

Scale of training data, not sensor resolution, delivered generalisation. Implanted BCI is now testing the same idea.

Method

Participants

Thousands of consenting participants for training; separate test users

Interface

Dry-electrode sEMG wristband

Signal

Surface electromyography at the wrist

Task

Continuous navigation, discrete gestures and handwriting

Decoder

Generic neural-network decoders, optionally personalised

Results

0.66 target acquisitions/s

Continuous navigation

0.88 detections/s

Discrete gestures

20.9 words/min

Handwriting

16%

Gain from personalisation

Handwriting models

Strength of evidencePeer reviewed; large participant pool; authored by the developer.

Figures are quoted from the published abstract and were checked on 3 Oct 2026.

Limitations

• Requires intact motor pathways to the wrist.
• Reads muscle activation, not brain activity.

What this could enable

Neuromotor input as an everyday interface, and a template for cross-subject decoding.