Products

Axonnic Uses an EMG Sleeve to Collect Robot Training Data

YC F26 startup Axonnic is developing a forearm sleeve that reads EMG signals, decodes finger movements, and estimates fingertip force. The aim is to give dexterous robots the exertion data that video capture cannot provide.

Font size

YC F26 company Axonnic has built a forearm sleeve that reads EMG signals, decodes finger movements, and estimates fingertip force. Its target is not VR, but training data for dexterous robots.

Axonnic is working on something that does not look much like consumer electronics: wrapping the forearm in a sleeve so the device can read muscle activity and decode finger movements and fingertip force from it.

It is not a motion-capture glove. The hand stays bare, while the sleeve records only the EMG signals produced by forearm muscles.

The team opened its Launch HN post with a straightforward introduction:

Hi HN! We’re Arsh and Max, founders of Axonnic (YC F26)

Axonnic is a YC F26 company founded by Arsh and Max. The current focus of the product is not connecting EMG to a VR game, but collecting human movement data to generate training data for dexterous robots.

The idea is worth breaking down. Video motion capture answers the question of “where is the hand,” while what robots often lack is “how much force did the hand apply?”

The sleeve measures force, not just movement

Axonnic’s forearm sleeve reads muscle activity, or EMG, to decode finger movement and estimate fingertip force.

This is fundamentally different from conventional optical hand tracking.

A camera sees the position, pose, and movement path of the hand. EMG sees the electrical activity that forearm muscles produce during specific movements. The former records the result; the latter attempts to capture part of the physiological process that produced it.

Consider a simple example:

  • A finger pinches a cup. A vision system knows that several fingers made contact with the cup.
  • For the same motion, a lighter or harder pinch may leave different EMG signals.
  • Tactile or force sensors can measure contact force directly, but they usually have to be attached to the fingers or the object.

Axonnic places the measurement on the forearm and keeps the hand bare. Users do not need to wear gloves or attach multiple sensors to every finger.

The phrase “estimate fingertip force” deserves some caution, however. The sleeve does not directly measure the contact point; it infers the output from muscle activity. Accuracy, latency, and repeatability cannot be judged from a Shorts demo alone.

The demo video is here:

https://youtube.com/shorts/xBrocTc9cfAWe

Video alone loses critical training signals

Axonnic argues that EMG is better suited to robot training than video alone because video cannot directly record muscle activity or the exertion behind a movement.

That argument has a practical basis.

Suppose a robot has to learn how to unscrew a nut. Ordinary video might provide:

  1. How the wrist rotates.
  2. How the fingers approach the nut.
  3. How the hand path changes after two fingers make contact.

But video has difficulty telling a model directly:

  • How much force to apply at the start.
  • Whether the output should decrease once the threads begin to catch.
  • Whether a person readjusts finger placement during a slip or increases grip force.
  • How much muscular contraction corresponds to the same visible movement.

In visual imitation learning, these missing variables are often described as an underdetermined action. The same hand trajectory can correspond to many different force distributions. A model that learns only the trajectory may reproduce movements that look correct without being physically correct.

EMG is not complete tactile sensing. It cannot tell you exactly which object a fingertip touches or directly measure contact pressure at each joint. It can, however, add a signal about muscular intent and exertion to visual data.

For dexterous robots, that signal may be worth more than another few millimeters of camera-positioning accuracy. A failed grasp often means the timing, grip force, or sequence of actions was wrong, not that the hand failed to reach the right position.

How the product data enters robot training

Axonnic’s product pipeline starts with human movement recording and converts those recordings into training data for dexterous robots.

It can be understood as a hybrid data pipeline:

EMG muscle activity
    ↓
Finger movement decoding + fingertip force estimation
    ↓
Alignment with human movement recordings
    ↓
Dexterous robot training data

The most valuable part is not simply an EMG classification model, but the mapping between EMG recordings and robot data.

A usable training sample has to answer several questions:

  • Which muscle was active, and when?
  • Which finger movement corresponds to the EMG change?
  • How are estimated force levels aligned with robot gripper forces?
  • Which EMG features transfer across different tasks?
  • How are noise, drift, and individual differences handled?

The last two are especially difficult.

Surface EMG typically contains motion crosstalk. When making a fist, other forearm muscles may also become active. Sensor placement, skin impedance, sweat, and how tightly the sleeve is worn can all shift the baseline and alter the waveform.

In other words, the same pinching motion may not produce the same signal for User A and User B. If the data cannot be calibrated for each user, collecting more may not produce a better model.

Robots also do not have a full set of forearm muscles. A training system eventually has to map human EMG intent to robot joint torques, gripper pressure, or tactile feedback rather than asking the robot to imitate the voltage of a muscle.

Axonnic is demonstrating the collection entry point. How the data is labeled, calibrated, and cleaned, along with whether it actually improves robot policy success rates, is the other part that determines the product’s value.

What Launch HN reveals about the product

The confirmed product details include:

  • Axonnic is a YC F26 company.
  • The team was founded by Arsh and Max.
  • The device takes the form of a forearm sleeve.
  • It reads EMG, decodes finger movement, and estimates fingertip force.
  • The hand remains bare.
  • The product records human movement and generates training data for dexterous robots.
  • The Launch HN post reached the Hacker News front page and included a Shorts demo video.

Original launch post:

https://news.ycombinator.com/item?id=50522776

Based on the available information, Axonnic looks more like a robot-data infrastructure company than an EMG hardware startup. The hardware acquires the signal, but the data-production pipeline behind it is the core product.

Its competition extends beyond other EMG sleeve vendors:

  • Cameras are inexpensive, cover a wide range of environments, and are well suited to collecting visual trajectories at scale.
  • Motion-capture gloves measure joint angles and fingertip contact directly, but they are more cumbersome to wear.
  • Fingertip force sensors are closer to real contact force, but usually require per-finger attachments.
  • EMG sleeves add muscular activity and exertion intent, at the cost of noisy signals and pronounced individual differences.

The most competitive combination may not involve EMG replacing the other approaches. It may involve placing vision, motion capture, and EMG on a shared timeline. Vision handles geometry, EMG provides intent and force cues, and force sensors or tactile data handle calibration.

Editorial assessment

The important part of Axonnic’s thesis is not that EMG will immediately become a standard input for robots. It is that the company has identified a specific gap: video shows how a hand moves, but not how much force it applies. This gap directly affects the quality of training for grasping, twisting, and similar tasks, and buyers can more easily evaluate the result through policy success rates.

Robot companies are more likely to purchase datasets and collection services than to depend on a single sleeve over the long term. If EMG only supports action classification, its commercial value is limited. If it can reliably provide force labels aligned with robot gripper forces, it has a path into the training pipeline.

Unlike earlier EMG demonstrations, the target here is not digital avatars or VR control but dexterous robots directly. The validation will be more demanding: without a measurable gain in success rates, even a clean waveform is just another sensor.

EMG crosstalk, cross-user calibration, real-time latency, and fingertip force-estimation error remain unconfirmed. The launch materials also include no benchmark. Three figures will be worth watching: error across repeated collections, transferability across users, and the increase in policy success rates after adding EMG.

Source

https://news.ycombinator.com/item?id=50522776

Source: https://news.ycombinator.com/item?id=50522776