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

In July 2026, researchers posted a graph neural network (GNN) on arXiv that recognizes hand gestures from surface electromyography (sEMG) signals in 48 milliseconds with 99% accuracy. That result beats every comparable method published to date for this class of task.

How the algorithm works

The system reads electrical activity from the forearm muscles through the eight electrodes of a Myo armband and turns the resulting muscle activation patterns into a graph. The graph captures more than signal amplitudes. It also encodes the spatial relationships between muscles, showing how certain muscle groups fire together for each gesture. A graph neural network then processes this graph to classify the gesture. This topological representation is what separates the approach from earlier methods, which worked with "flat" sEMG time series.

Key system specs:

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

Why 48 ms matters

Latency is critical for controlling a prosthetic hand or an AR interface. People start to notice lag once it exceeds 50–100 ms, and that makes a device awkward to use or even uncontrollable. Previous ML approaches to sEMG often needed 200–300 ms, which is enough to make every movement feel sluggish and unnatural.

The key point is that the team hit 48 ms on an ordinary consumer CPU with no dedicated neural accelerator. That means the algorithm could potentially run directly on a chip inside a prosthesis or AR headset, with no cloud connection and none of the network latency that comes with one. Offline inference is what separates a research demo from a real medical device.

Where the method falls short

A sample of 8 subjects is typical for academic neural interface research, but it is too small to support confident generalization. All participants were healthy volunteers. Nobody has yet tested how the algorithm performs for people with altered muscle activity, such as actual prosthesis users.

The Myo armband also needs precise positioning every time someone puts it on, and a small shift in the electrodes can change the signal patterns. The authors do not address how robust the model is to this kind of variability, which leaves the question open for future work.

What it means

99% accuracy at 48 ms latency on an ordinary CPU is a bid to move beyond the academic stage. If the results hold up on a larger and more diverse sample, the GNN approach to sEMG could bring prosthetic and AR interface control closer to the speed and fluidity of natural hand movement, without expensive implanted electrodes.

Frequently asked questions

Why use graphs for sEMG instead of conventional neural networks?

A graph explicitly encodes the spatial relationships between muscles: which muscle groups activate together and in what order. Conventional convolutional or recurrent networks process each electrode channel independently and lose this topological information. A GNN accounts for it directly, which yields a richer representation of the gesture. According to the authors, this lets it outperform LSTM and CNN models on the same data.

What hardware was the 48 ms latency measured on?

The measurements were taken on an Apple M1 Pro, a mobile chip with no dedicated neural accelerator for this task. The authors cite this as evidence that the method suits embedded systems with comparable compute power.