Axonnic's EMG Sleeve Reports 6.9° Finger Error After Four Days, and Its Pages Disagree on 32 or 64 Channels
Hardware / news
Axonnic's EMG Sleeve Reports 6.9° Finger Error After Four Days, and Its Pages Disagree on 32 or 64 Channels
The Y Combinator startup sells robot-training sleeves from $8,000 each; every accuracy figure is its own, and only one comparison uses an outside dataset.

Axonnic, a Y Combinator-backed startup out of Oxford, is selling a forearm sleeve that reads muscle signals to estimate finger angles and fingertip force, and its own pages give two different channel counts for the same hardware: 64 in its Launch HN post, 32 on its homepage.
The company's Launch HN post tagged YC F26 had drawn 55 points when the lead sheet pulled it. The product records skilled workers doing manual tasks, so that the data can train dexterous robots. The robots never wear the sleeve.

Price and what is shipping
Axonnic's homepage lists a Gen 2 preorder for U.S. shipping only. A single sleeve is $12,000 with a $2,400 deposit, pairs are $10,000 per sleeve, and orders of 5 to 99 pairs are $8,000 per sleeve. Gen 1 is at full capacity. Deposits are refundable until the order ships.
Customers are thin. A European manufacturer has signed a letter of intent for robotic sewing support, unnamed, and a construction pilot is recording electricians, plumbers and demolition crews. The page lists no named commercial customer and discloses no funding amount. Backers listed are Y Combinator, Boost VC, Sterling Road and Oxford Seed Fund. The founders are CEO Maxim Williams and CTO Arsh Patankar.
The accuracy figures, and the conditions
Every number below is published by Axonnic, not by an independent party, and none has been peer reviewed.
| Claim | Figure | Condition |
|---|---|---|
| Finger error, homepage headline | 4.58° | Held-out recordings, Vision Pro reference |
| Four days after training | 6.9° | 215 seconds, no recalibration |
| Assembly session, participant 1 | 5.3° | Two participants only |
| Assembly session, participant 2 | 6.5° | Two participants only |
The reference in each case is Apple Vision Pro hand tracking, which the benchmarks page quotes Apple as calling "imperfect, especially during heavy occlusion." That means the sleeve is scored against another estimator rather than ground-truth motion capture. The sample is small: two people in the assembly session, ten sessions in the band comparison.
Against Meta's published model
The one independent yardstick is Meta's emg2pose benchmark. The paper on arXiv describes 193 users, 370 hours and 29 stages, with labels from a 26-camera motion-capture rig and a 16-channel wrist sensor sampled at 2 kHz. Axonnic's benchmarks page says it re-trained Meta's code, got 12.25° against the paper's 12.2°, and then swapped in its own model on Meta's data.
- Axonnic model11.3 degrees
- Meta baseline12.2 degrees
Source: Axonnic benchmarks page (vendor-supplied), accessed 2026-11-18
That is 7.7% lower error on unseen people, 8.4% on unseen movements and 9.5% on both. Axonnic reports lower error for all 20 held-out people. Only the model changed, not the hardware, so this tests software on Meta's 16-channel data and says nothing about the sleeve itself.
What the pages disagree on
Besides the channel count, the benchmarks page cites 12.0° on index knuckle flexion for unseen people against 4.58° on the homepage, which measure different things. It also shows that wrist contacts died in most sessions, and that force estimates need per-user calibration: 2.0 N error for participant 1 and 3.5 N for participant 2.
On Hacker News, commenters raised signal drift over an hour of gripping, and a founder said drift tests are planned. Others pointed to Meta's EMG wearable work, and several objected to wearing tracking devices at work.
Why robot-hand data is being bought
The premise is that robots learn dexterity from human demonstrations. Genesis AI's $500 million round also centres on robot hands, and the site's report that Tesla's Optimus line in Fremont has no published output number is a reminder of how little hand-level production data is public.
What would change this read
A third-party run on a task the company did not choose would change it, as would a named customer paying beyond a letter of intent. The next milestone is shipment of Gen 2 units; the page gives no date.
Sources
More in Hardware
- 01Starship's Ship 42 Is Caught After 14 Orbits, and Two Reports Give Different Reasons the Booster Was NotFlight 15 returned an upper stage to the tower for the first time and dropped 26 Starlink V3 satellites; SpaceX has no reflight date in these accounts.
- 02Genesis AI Raises $500 Million at $3.5 Billion, and Its CEO Says Eno's Hands Last WeeksThe Paris robot maker has built 30 to 40 Eno units and chose a US contract manufacturer, with no price and no paying customer count.
- 03Tesla Cybercab's Eight NHTSA Crashes Include a 53 mph Highway Hit With a Test Driver AboardThe four driverless crashes stayed under 5 mph; the Robotaxi median pre-crash speed is 1 mph, which says more about the geofence than the software.
- 04Anthropic's Three-Day AMD Port Covered Prefill for One Model on MI355XThe roadshow slide credits Claude Mythos Preview with a six-to-12-month job; it leaves out decode, the kernel denominator and any Nvidia comparison.