Axonnic's forearm sleeve reads muscle signals to decode finger motion and estimate fingertip force, leaving hands bare
Synopsis
Axonnic released a forearm sleeve that records muscle electrical activity with 64 channels of surface electromyography to decode finger joint angles and estimate fingertip force while the hands stay bare. The team argues video captures movement but not the effort behind it, and that barely visible muscle activity, such as holding a grip, is the key signal for sewing and delicate assembly. In one 215-second recording made four days after training without recalibration, the sleeve decoded 20 joint angles with 6.9° error, compared against Apple Vision Pro hand tracking.
Interpretation
Axonnic's forearm sleeve leaves the hand uncovered and reads electrical activity from forearm muscles through electrodes on the skin, decoding finger motion and estimating fingertip force to collect human movement data for dexterous robots. The stated comparison is training robots from video alone: video captures movement but not muscle activity or the effort behind it, while a useful signal can be present even when the hand barely moves, since maintaining a grip itself involves muscle activation. In a 215-second recording made four days after training, the sleeve decoded 20 joint angles without recalibration at 6.9° error, compared with Apple Vision Pro hand tracking.
Sixty-four channels of surface EMG are distributed across the forearm so the system does not overfit to one muscle group; EMG is synchronized with hand tracking to train a model that predicts finger joint angles, and the decoder then runs solely from EMG. Fingertip force estimates are added from calibration against finger presses on force sensors, handled separately from joint-angle decoding. Because taking the sleeve off and putting it back on changes electrode positions, tests are run over separate wearings, and the 215-second recording is one such wearing.
Perspective
What is shown is recording and data collection with the forearm sleeve in industrial settings: sewing and other manual tasks at a European manufacturer that signed an LOI for support with robotic implementation, demonstration capture involving electricians, plumbers and demolition work, and a public demo of controlling a Proception robotic hand from forearm EMG; these are pilot and data-collection examples.
The team reports wireless interference and unreliable off-the-shelf hardware during factory recordings, which produced some bad recordings and required repeating data collection, and is therefore developing its own electronics with a wired PCB design to eliminate interference and timing inconsistencies; the joint-angle error result comes from one wearing and one 215-second recording.
