Axonnic's YC F26 EMG Sleeve Claims 58% Fresh-Day Tracking, From Its Own Benchmarks
Startups / news
Axonnic's YC F26 EMG Sleeve Claims 58% Fresh-Day Tracking, From Its Own Benchmarks
The robot-training sleeve beat Meta's published emg2pose model by 0.9 degrees on one split, but the force and long-term results rest on two wearers and one clip.

Axonnic, a Y Combinator F26 company, launched a 64-channel forearm sleeve on Hacker News that reads muscle signals to estimate finger movement and fingertip force, for recording human skill to train dexterous robots. Every accuracy figure comes from the company's own benchmarks page, and the post discloses no funding amount or price.
The founders, who give only the first names Arsh and Max, say the work grew out of Max's research on high-density surface EMG at Oxford. The sleeve leaves the hand bare, which is the pitch against glove-based capture.
The vendor-supplied numbers
The benchmarks page uses Apple Vision Pro hand tracking as the reference for joint angles. Apple calls its own hand labels imperfect during heavy occlusion, so the yardstick has error of its own.
| Test | Axonnic result | Comparison |
|---|---|---|
| Fresh-day tracking, 10 sessions | 58% | Single bands 4% to 44% |
| Fingertip presses caught in 0.5 s | 84% | Pressure glove reference |
| Free-grasp force, R² | 0.81 | Best single band 0.50 |
| Four days, no recalibration | 6.9° error | One 215-second clip |
The fresh-day test trains on earlier sessions and scores a later wearing after the sleeve is removed and refitted, which is the hard case for EMG. A wrist band alone fell to 15 percent on re-wear while the sleeve held 51 percent.
The Meta comparison
The company also re-ran Meta's model on Meta's 193-person emg2pose dataset, using Meta's evaluation code. Its retrained copy of Meta's model scored 12.25 degrees against the 12.2 degrees in Meta's paper.
| Split (average error) | Axonnic | Meta |
|---|---|---|
| People unseen | 11.3° | 12.2° |
| Movements unseen | 13.9° | 15.2° |
| Both unseen | 14.3° | 15.8° |
The page says only the model changed, and that Axonnic's came out ahead on every split and for all 20 held-out people. The gaps are 0.9, 1.3 and 1.5 degrees, and the page reports no confidence intervals or variance. Nothing on it is an independent replication.

Where the sample is thin
The November assembly session used two participants. Joint-angle error was 5.3 degrees for the first and 6.5 for the second, with per-task results from 4.6 degrees on screw caps to 8.1 degrees on fabric and zips. Force error was 2.0 newtons for the first participant and 3.5 for the second, so the second wearer's force error was 75 percent higher.
The four-day drift result is one clip. On Hacker News, a commenter named millspore asked how the model copes with EMG baseline drift from fatigue. Co-founder maxsw345 replied that readjusting the grip helps and that the team would test drift. The company says about two minutes of ordinary movement recalibrates the sleeve.
Customers and the vendor-benchmark problem
The post says a European manufacturer is using the sleeve to record sewing and other manual tasks and has signed a letter of intent for robotic implementation support. Construction pilots covering electricians, plumbers and demolition work are described as data collection. A public demo drives a Proception robotic hand from forearm EMG. An LOI is not a purchase order, and no revenue is given.
Self-published benchmarks deserve the same discount applied to OpenComputer's 3.3x Devin claim, which rested on one task in four repositories. The hardware question is also familiar: Agility's Digit 5 launch gave runtime and weight but no unit count, and Axonnic gives no count of sleeves built.
The company is building a wired PCB to cut wireless interference and timing problems seen in factories. A third-party replication on the emg2pose split, or a named customer's data-collection volume, would be the first outside check.
Sources
More in Startups
- 01Paytm Pays Rs 162 Crore for 19.1% of CardUp, Which Implies Rs 848 Crore, and Its Filing Dates DisagreeThe Singapore payments firm's turnover rose to Rs 299.2 crore in 2025, so the minority stake prices it at under three times that figure.
- 02Humyn Labs' $47 Million Round Is $26 Million Equity, and Its Revenue Claims Differ by $40 MillionThe Bengaluru robotics-data startup says it passed a $50 million annualised run rate, while the figure for revenue actually booked from April to October is $10 million.
- 03Payhawk's $280 Million Series C Doubles Its Valuation to $2 BillionGoldman Sachs Alternatives led the announced round, which is subject to regulatory approval, and Payhawk disclosed no revenue to set the price against.
- 04Bending Spoons Buys Mailchimp for $3.68 Billion, and Its Two Filings Disagree on 'All-Cash'Intuit sold the email marketing business for about 69 percent less than the roughly $12 billion it paid in 2021, and the press release and the 6-K describe the payment differently.