Benchmarks · UPDATED 4 NOV 2026

Our sleeve beats a band.Our model beats Meta’s.

We built a 64-channel sleeve because one band is a gamble. Put a single band back on the next day and it can lose the hand almost completely: one of ours tracked 4 percent of the movement. The full sleeve tracked 58. Then we put our model on Meta’s own benchmark, with Meta’s data and Meta’s scoring, and it beat Meta’s for all 20 people they held back.

of the movement one single band tracked on a fresh day. Our sleeve: 58%.^⁠a: 4%
of fingertip presses caught within half a second, read from the forearm. Motion sensors alone: 4%.^⁠b: 84%
Meta’s held-out people got a lower error from our model^⁠c: 20 of 20

One band is agamble.

A single band has to land in the right spot and keep every contact, every time it goes on. That’s where bands fall over. We trained on earlier sessions, then tested on a fresh day, after the sleeve came off and went back on. The full sleeve still tracked 58 percent of the hand’s movement. Single bands managed 4 to 44 percent. One position tracked almost nothing.

Contacts die in the real world. Two did at the wrist in most of our sessions, and the wrist band carrying them fell to 15 percent across our re-wear tests. The sleeve wore the same dead contacts and still tracked 51 percent. Sixty-four channels never rely on one spot.

Even in the easiest test, the same session the model learned from, every band had 29 to 39 percent more error than the sleeve.

Movement tracked = 1 − error ÷ error of a hand held in one average pose. Fresh-day test: trained on earlier sessions, tested on a later wearing. 10 sessions, error against Apple Vision Pro hand tracking, same model and settings for every band.

Force a cameracan't see.

A finger resting on a cup and a finger crushing it look the same on video. The sleeve reads the muscles doing the pressing.

Against a pressure glove, it catches 84 percent of fingertip presses within half a second. On free grasps it tracks fingertip force with an R² of 0.81, where the best single band managed 0.50. Motion sensors on the arm alone caught 4 percent of presses.

The glove we used to check it makes our point for us. Gloves cost dexterity: pegboard scores drop 9 to 29 percent with gloves on,^⁠[4]^⁠[5] and a single latex layer dulls touch by 17 percent.^⁠[6] The sleeve leaves the hand bare.

Middle fingertip, one held-out stretch, relative units.

Scored on held-out data against a Proception ProGlove pressure glove.

No line of sightneeded.

Headsets track hands with cameras, so they lose them in a grip. In the most careful public test, a Quest 2's knuckle error rose from 11.5° to 16.7° just from moving faster.^⁠[2] Apple calls its own Vision Pro hand labels “imperfect, especially during heavy occlusion”.^⁠[3]

From forearm muscle alone, on people it has never seen, our model averages 12.0° at that same knuckle. Camera-grade, without a camera.^⁠[1] Muscle signals don't care what is in front of the hand.

Camera error grows with speed^⁠[2]

Ours: index knuckle flexion, 20 held-out emg2pose people, all movements, against 26-camera motion capture.^⁠[1] Quest 2: 8 people opening and closing the hand, against 12-camera motion capture.^⁠[2] Different studies and tasks.

And when the headset does lose the hand, the sleeve covers for it. Over one to ten seconds of lost tracking, sleeve plus headset stays within 5.9° to 7.5°. The headset's frozen last frame drifts to 10° to 16°.

Gaps simulated on a held-out recording by hiding the headset's frames, then scored against them.

THE MODEL

Same data.Same test.A better model.

This one is pure software. Meta recorded 193 people on its own 16-channel wristband and judged the results against a 26-camera motion-capture rig. We took those same recordings, trained on the same people, tested on the same 20 people Meta held back, and scored with Meta's code. Same hardware, same data. Only the model changed, and ours came out ahead on every split.

LOWER IS BETTER · MODEL VS MODEL, SAME DATA

People the model has never seen: 11.3°; Meta 12.2°
Movements it has never seen: 13.9°; Meta 15.2°
Both at once: 14.3°; Meta 15.8°

Same test, every split

Average error across 20 finger joint angles, measured against motion capture. The axis starts at 10° so the gaps can be read. Meta's figures are from Table 4 of their paper.^⁠[1]

Twenty out oftwenty.

An average can hide a few bad results, so we counted people instead. Meta kept 20 people out of training. Every one of them got a lower error from our model than from Meta's.

To make sure the test was fair, we first re-trained Meta's model with Meta's code. It scored 12.25°. Their paper says 12.2°. Then we ran ours on exactly the same test.

It holds in millimetres too: 14.4 mm average landmark error against Meta's 15.8 mm.^⁠[1]

Each line is a person Meta held out of training. Lower is better.

Four dayslater.

We trained a decoder, took the sleeve off, and came back four days later. With no recalibration it still tracked the headset at 6.9° average error and a 0.87 correlation. If the sleeve shifts, two minutes of ordinary movement recalibrates it.

Average error, no recalibration: 6.9°
Correlation with the headset: 0.87
To recalibrate if the sleeve moves: 2 min

Nov 2026 session

4.6° error on screw caps,from the forearm alone

We recorded a full session of real assembly work — connectors, cords, caps, fabric and delicate parts — and scored the model on segments it never trained on. Mean joint-angle error came in at 5.3° for participant 1.

mean joint-angle error, held out (participant 1): 5.3°
mean joint-angle error, held out (participant 2): 6.5°
fingertip force error at r = 0.91, from about 2 minutes of training data: 2.0 N

Held-out joint-angle error by task

Degrees, lower is better. Each bar is scored on 30 s segments excluded from training.

Fingertip force fromtwo minutes of calibration

After about two minutes of presses against a gauge, the sleeve estimates fingertip force within 2.0 N on average (r = 0.91). On our second participant it's 3.5 N (r = 0.84). The same model then reads force continuously through the tasks.

One sleeve, one session, training-ready manipulation data.

Reserve yours

Held out means scored on 30 s segments excluded from that model's training, within the same session. Force is scored on held-out gauge presses.

How we measured

Sleeve vs band^a	10 sessions, Vision Pro reference, same model and settings for every band; fresh-day test = trained on earlier sessions, tested on a later wearing.
Force^b	Held-out data, ProGlove pressure-glove reference.
Knuckle	index MCP flexion, emg2pose held-out people^⁠[1]; Quest 2, 8 people, 12-camera Qualisys.^⁠[2]
Headset gaps	Simulated gaps on a held-out recording.
Meta's benchmark^c	emg2pose test split, 20 held-out people, 26-camera OptiTrack; our model scored with Meta’s evaluation code; Meta = Table 4.^⁠[1]
Four days later	215-second clip, Vision Pro reference, 20 joint angles.
Nov 2026 session	Pose: each wearing cut into 30-second blocks, five-fold; every block scored by the fold model that never trained on it; Vision Pro reference, 20 joint angles. Force: five-fold within the gauge-press recording; gauge reference read from video.

References

Want the raw numbers?

We'll walk you through the full benchmark pack.

Start a pilot