- Author
- Patrick Kaifosh, Thomas R. Reardon and CTRL-labs at Reality Labs
- Published
- 23 July 2025
- Kind
- Peer-reviewed open-access journal paper, with tests on new users
- Original
- www.nature.com/articles/s41586-025-09255-w
- Subject
- Science and medicine
What it is
CTRL-labs at Reality Labs wrote this open-access paper. The first two authors are Patrick Kaifosh and Thomas Reardon. It sets out to build a computer input that needs no calibration. The device reads electrical signals from muscles at the wrist. Tasks were cursor control, gesture input and handwriting.
What it says
- The paper describes a wristband with 16 bipolar channels that records muscle signals.
- It says models trained on thousands of people work for new users without calibration.
- Median closed-loop speeds were 0.66 and 0.88 acquisitions per second, and 20.9 words per minute.
- Models trained on one person did poorly when tested on another person.
- Error fell as training participants grew, following a power law.
- The sEMG decoders were slower than a trackpad, a game controller and open-loop handwriting.
- Offline, personalizing a handwriting model gave a 16% median gain from 20 minutes of data.
What evidence it gives
Training data came from 162 to 6,627 participants, depending on the task. Closed-loop tests used 17, 24 and 20 new participants. Baselines were a trackpad, a Joy-Con controller and open-loop handwriting. Code and recordings from 300 participants are released.
Interests and limits
The authors are from CTRL-labs at Reality Labs. Data were collected at Meta facilities or vendor sites. The decoders were slower than every baseline tested. Handwriting needed coaching for about 80% of test participants. Training used able-bodied participants, and the paper says clinical use is unclear. Figures and videos were not viewed.