ExoHand: A £200 Closed-Loop Hand Exoskeleton, Calibrated in 22 Seconds

Our aim is to show a complete closed-loop hand rehabilitation system working in real time on commodity hardware, and to be concrete about how little data the decoder behind it needs. ExoHand pairs a 3D-printed tendon-driven exoskeleton with live surface electromyography (sEMG) decoding for stroke patients who have lost hand function. The decoder is fit from scratch at the table, on 22 seconds of the user's own signals, with no pretrained model in the loop.

One participant wears the exoskeleton. Two servos sit on a forearm cuff, one driving the four fingers and one the thumb so that thumb motion does not interfere with the rest, each geared 2:1 to give the winches a full 360 degrees of travel. Every finger has its own winch, sized so its circumference matches the arc that fingertip sweeps between closed and open, which lets fingers of different lengths arrive in position together. Flexion and extension tendons run from the same winch, one routed over a dorsal plate to the fingertips and the other around to the palmar side, so turning the winch reels in one while releasing the other and the hand stays under controlled tension throughout.

A second participant wears a forearm sleeve holding four sEMG sensors over FCR, ECR, FDS and EDC, and we calibrate a HistGradientBoosting classifier to them during a guided two-minute session. The model then decodes their intent live: when they close their hand, the exoskeleton closes the first participant's hand; when they open, it opens. Attendees swap roles, so each person drives the device and feels it move their fingers. Clinically the loop runs within one patient, completing the movement their own pathway cannot; we separate it here so observers can watch the decode travel between two people rather than infer it from a screen.

The relevance to this workshop is the closed loop itself, and what we measured while building it. Before settling on per-user calibration we tested transfer from a large healthy-population corpus, GrabMyo, at 1.14 million windows across 43 subjects, into stroke intent decoding on 48 patients. Across five tests we found no transfer we could detect: zero-shot accuracy sat at 0.360 against a 0.333 chance floor, while 22 seconds of on-body data reached 0.896. Those offline results used a four-channel per-patient subselection matched to the rig we bring, so the hardware on the table and the analysis behind it see the same channel count.

Acquisition, 20 Hz envelope extraction, inference and actuation run on a Teensy 4.0, a laptop, hobby servos and off-the-shelf sensors, for a total build cost under £200. Nothing is pre-recorded and nothing touches a network. The live interface shows the four EMG channels, the decoded hand state, and a rep counter with session accuracy and stability, so the path from signal to movement stays legible to everyone at the table.