Graph neural network recognizes hand gestures in 48 ms: 99% accuracy for prosthetics and AR

Researchers presented in July 2026 on arXiv a graph neural network (GNN) that recognizes hand gestures from surface electromyography (sEMG) signals in 48 milliseconds with 99% accuracy — a result exceeding all previously published analogs in this class of tasks.

How the algorithm works

The system reads the electrical activity of forearm muscles through eight electrodes of the Myo armband and converts muscle activation patterns into a graph. The graph encodes not just the amplitudes of signals, but also spatial relationships between muscles — how certain muscle groups are activated together with others for each gesture. This graph is then processed through a graph neural network for classification. This topological representation distinguishes the approach from predecessors that worked with "flat" sEMG time series.

Key system characteristics:

• Sensor — Myo armband with 8 electrodes around the forearm
• Architecture — graph neural network (GNN)
• Latency — 48 ms (graph construction + prediction) on Apple M1 Pro
• Classification accuracy — 99% (average across 8 subjects)
• Test participants — 8 healthy volunteers

Why 48 ms is an important number

For controlling a prosthetic hand or AR interface, latency is critical: humans begin noticing "desynchronization" already at delays over 50–100 ms, and this makes the device uncomfortable or unmanageable. Previous ML solutions for sEMG often required 200–300 ms — enough to make each movement feel sluggish and unnatural.

Key in the result: 48 ms was achieved on a standard consumer CPU without specialized neural accelerators. This means the algorithm could potentially run directly on a chip inside a prosthetic or AR headset — without cloud connectivity and associated network delays. It is precisely the possibility of offline inference that makes the difference between a research demo and a real medical device.

Where the method's limitations lie

A sample of 8 subjects is typical for academic work in the neurotechnology field, but small for confident generalization. All participants were healthy volunteers: how the algorithm will perform in people with altered muscle activity — for example, actual prosthetic users — has not yet been tested.

Additionally, the Myo armband requires precise positioning each time it is put on: even a small shift in electrodes can change signal patterns. The authors do not address the model's robustness to such variability, which remains open for future work.

What this means

99% accuracy with 48 ms latency on a standard CPU — this is a claim to exit the academic stage. If results are reproduced on a larger and more diverse sample, the GNN approach to sEMG could bring prosthetic control and AR interfaces closer to the speed and smoothness of natural hand movement — without costly implanted electrodes.

Frequently asked questions

Why use graphs for sEMG instead of ordinary neural networks?

A graph explicitly encodes spatial relationships between muscles: which muscle groups activate together and in what order. Ordinary convolutional or recurrent networks process each electrode channel independently, losing this topological information. GNN accounts for it directly, which provides a richer representation of the gesture and, according to the authors, allows surpassing LSTM and CNN on the same data.

On what equipment was the 48 ms latency measured?

Measurements were conducted on the Apple M1 Pro processor — a mobile chip without a dedicated neural accelerator for this task. The authors present this as an argument in favor of the method's suitability for embedded systems with comparable computational power.