Open, Accurate, and Calibration-Free Muscle-Computer Interfaces
Ethan Eddy, Evan Campbell, Erik J. Scheme, Scott Bateman · 2026 · Proceedings of the 2026 CHI Conference on Human Factors in Computing Systems (CHI '26) · doi:10.1145/3772318.3790689
Summary
Eddy and colleagues tackle one of the long-standing barriers to practical muscle-computer interfaces (MCIs): that accurate, cross-user gesture recognition from forearm EMG has historically required either per-user calibration or access to closed datasets and proprietary hardware. Building on a 2025 milestone from Meta / CTRL-Labs, which showed that a large-scale dataset plus deep learning can produce calibration-free MCIs but only via closed resources, this paper re-implements that vision entirely in the open. The authors train foundational models on the publicly available EMG-EPN-612 dataset (612 participants, Myo Armband, 8 channels at 200 Hz) using the open-source LibEMG library. They release two calibration-free models: a continuous model for 1D cursor control via wrist flexion / extension, and a discrete model for five-class gesture recognition (rest, hand open, hand close, wrist flexion, wrist extension, double-tap). Twenty new participants tested the cursor interface and fifteen tested the discrete interface with no calibration data collected from any of them. Reference baselines were collected using a trackpad (Task 1) and a keyboard (Task 2). The contribution is deliberately reproducibility-oriented: all code, data pipelines, and pre-trained weights are released on GitHub, making this the first time researchers outside Meta can independently replicate, benchmark, and extend calibration-free MCI results.
Key findings
For 1D cursor control, the open calibration-free MCI achieved a mean target acquisition time of 1.15 s by the final block — comparable to the 1.52 s Meta reported for the same task on closed proprietary hardware, and within about 2x of the trackpad baseline (0.58 s). Only two of twenty participants performed worse than Meta's reported baseline. For discrete gesture recognition, the system reached a final response time of about 1.0 s and an error rate of 2% on the first try (dropping to 0.5% on the second try) — approaching the keyboard baseline (1.02 s response, 1.3% error). Participants reached proficiency after roughly 25-50 trials (2-3 minutes), showing a significant learning effect across blocks despite the model being calibration-free. Across both tasks, participants used relatively low contraction intensity, suggesting muscle fatigue was not a major concern. Three participants with arthritis, hand tremors, or tennis elbow completed the tasks without difficulty. The paper demonstrates for the first time in an open setting that data collected five years earlier in a different country generalises to new users in closed-loop real-time control.
Relevance
For accessibility practitioners, this paper is a concrete step toward wearable, calibration-free muscle-based input that could serve people with severe physical disabilities who cannot reliably operate keyboards, mice, or touchscreens. Because EMG picks up electrical activity from muscle contractions directly, it can register subtle gestures from users with very limited range of motion, including individuals with high spinal-cord injuries, ALS, or late-stage muscular dystrophy for whom even small residual muscle activity remains. Crucially, by open-sourcing the models, data pipeline, and evaluation harness, the authors lower the barrier for AT developers, clinicians, and researchers to prototype custom gesture sets and integrations with AAC, cursor control, or switch-style access without proprietary SDKs. Caveats for AT work: the Myo Armband has been discontinued (though the open models may transfer to successors like OyMotion, MindRove, and SifiLabs); the evaluation was in controlled seated conditions so false activations during daily living remain an open problem; and the gesture sets evaluated assume intact forearm anatomy, so application to users with limb differences would require additional data and validation.
Tags: muscle-computer interface · electromyography · EMG · gesture recognition · machine learning · wearable technology · alternative input · myoelectric control · open source · prosthetics