Technology

Axonnic's Forearm Sleeve Reads Finger Movements and Grip Force to Train Robots

Martin HollowayPublished 12m ago5 min readBased on 2 sources
Reading level
Axonnic's Forearm Sleeve Reads Finger Movements and Grip Force to Train Robots
source:axonnic.com

Axonnic, a startup founded by Max and Arsh, says its wearable forearm sleeve tracks individual finger movements and fingertip pressure in real time, using electrical signals from the forearm. It needs no cameras and no gloves. The company is pitching the device to manufacturers as a way to automate dexterous tasks: record their best workers, then train robots on that data.

The sleeve is an electromyography (EMG) device. EMG sensors pick up the electrical signals that muscles produce when they contract, so forearm activity can be read without covering or touching the hands. Axonnic says it estimates fingertip force while leaving the hands completely free. It presents this as capturing information that cameras miss, such as how hard a grip is applied rather than where the fingers sit. Axonnic positions the sleeve against camera-based hand tracking and instrumented gloves, both of which it considers less practical on a working floor.

On its finger-tracking model, Axonnic reports a 4.58-degree average joint-angle error at 50 predictions per second. The model was trained on 78 minutes of data. The company states it was validated against Apple Vision Pro hand tracking on held-out recordings, meaning data the model never saw during training. It also says that after a wearer takes the sleeve off and puts it back on, two minutes of movement re-calibrate the model in under a minute.

The most demanding figure concerns session-to-session stability. In a headset-versus-sleeve test, Axonnic reports a 6.9-degree mean joint-angle error and a 0.87 correlation with the headset, where 1.0 would be a perfect match. The decoder in that test had been trained four days earlier on a separate wearing of the sleeve and was never recalibrated.

On the robotics side, Axonnic claims its full-sleeve EMG signals and model drive a robotic hand the way a human hand moves, in real time and without cameras. The company describes its data pipeline as follows. Workers wear the sleeve while being filmed with hand tracking. The EMG, video and hand-pose streams are aligned into a single dataset, and the demonstrations are retargeted, meaning mapped onto the robot's own range of motion, to train a policy: the learned set of rules that decides what the robot does next.

Axonnic says its sleeves are shipping and that paid pilots are underway with its first design partners. It names three target tasks: guiding fabric through a sewing machine, handling an electrician's tools, and positioning small parts during delicate assembly. These examples share a property the company appears to be targeting. They depend on continuous force and grip judgment that is hard to write in code or capture with a fixed camera rig. The company lists Boost VC and Oxford Seed Fund as backers. Max studied Biology (MBiol) at Oxford, and Arsh studied medicine there.

Most of the figures above describe controlled conditions: a held-out dataset, a single wearer, a four-day interval. The gap between a 6.9-degree error in a test and a stable grip on a shop floor across shifts, sweat, sleeve placement and different operators is where the commercial case will be decided. The pilots are the first place that gap will show.

In my view, the more interesting question is what the approach does to the economics of automation. Most robot training today relies on teleoperation rigs, scripted demonstrations or expensive simulation. A method that records a skilled operator's hand behavior while that operator works normally would move the bottleneck from engineering time to access to experienced workers. It would also put a wearable sensor on the people whose technique is being captured. Both points deserve scrutiny from manufacturers and workers alike.

The broader context here is that the same sensing stack could make expert craft transferable to machines that previously needed a human in the loop for fine manipulation. If the recalibration and cross-session figures hold outside the lab, the sleeve could reduce the setup cost of teaching robots delicate tasks. That is the outcome Axonnic is betting on.