Apple trained artificial intelligence to recognize unfamiliar handshakes based on EMG signals

Apple has created the EMBridge model – gesture recognition from EMG signals

New research by Apple shows that their AI, EMBridge, can determine hand movements solely from electrical muscle activity (EMG) signals, even when such gestures were not present in the training set.

What is EMG and where it’s already used
Electromyography measures the electrical activity generated during muscle contraction.

In medicine it is used for diagnosis and physiotherapy, as well as in limb prostheses.

Wearable devices (for example, Meta Ray‑Ban Display glasses with a Neural Band controller) use EMG to control virtual reality.

How EMBridge was trained
1. Data – researchers used two open datasets:

* `emg2pose` – EMG signals and hand coordinates.

* `NinaPro DB2` – a similar dataset.

2. Two views – the model was first trained on two separate streams:

* only EMG signals;

* only hand position data.

3. Synchronization – after initial training, researchers “connected” the streams: the component working with EMG learned to “understand” information from the coordinate data. As a result, EMBridge could recognize gestures using only the EMG signal.

Increasing the difficulty of the task
* A portion of the second stream (coordinates) was cut and the model was forced to make inferences solely on EMG.

* To avoid excessive errors, prediction evaluation became less strict: similar gestures were treated as related rather than entirely distinct.

* This approach helped “structure” the feature space and improved reconstruction of hand positions that were not present during training.

Testing and results
* The model was tested on the same `emg2pose` and `NinaPro` sets, using them as benchmarks.

* EMBridge maintains high accuracy even when using only 40 % of the training data.

Limitations
Scientists emphasize that a key obstacle remains access to datasets pairing “EMG + hand position.” Such data are still limited in volume and not always available.