ZICQ
中 Log in / Sign up
ZICQ Info Agentic #Axonnic #Surface EMG #Robot Learning #Human-Machine Interface #Dexterous Manipulation

Axonnic Uses 64-Channel EMG to Capture Dexterity Training Data for Robots

Avatar of Mr.Xu

By Mr.Xu Community Post

Published:

中文阅读 (Chinese) English Version
Co-reading Select any sentence in the article to highlight, annotate, or copy a quote card. Join discussion ↓

Summary:Axonnic has launched a forearm sleeve that uses 64-channel surface electromyography to decode finger motion and estimate fingertip force while leaving the hand uncovered. During training, EMG is synchronized with hand tracking to create joint-angle labels; the reported inference path relies on EMG alone, with force calibrated through presses on force sensors. In one reported session, a 215-second recording made four days after training without recalibration produced a 6.9° aggregate error acros


Core Facts

Axonnic has publicly demonstrated a forearm surface-electromyography system. Its sleeve uses 64 EMG channels distributed across the forearm to decode 20 finger-joint angles and estimate fingertip contact force while leaving the hand uncovered. The system is intended not to replace every vision-based approach, but to enrich robot demonstrations with muscle activation, effort intent and force information that video cannot directly expose.

The disclosed training pipeline synchronizes EMG with hand tracking and uses the tracker to generate joint-angle labels. After training, the reported inference path uses EMG alone. A separate force estimate is calibrated through finger presses against a force sensor. Axonnic has also demonstrated forearm-EMG control of a Proception robotic hand and collected pilot data from sewing, electrical, plumbing and demolition tasks. A European manufacturer has signed a letter of intent supporting robotic implementation, but these deployments remain pilot and data-collection exercises.

Wireless interference and inconsistent timing from off-the-shelf hardware caused failed factory recordings and repeated data collection. The team is therefore developing its own wired PCB to improve electrical isolation and timing consistency.

Technical Mechanism: From Muscle Activity to Joint Angles and Force

1. Surface EMG exposes intent and effort cues

Surface EMG measures the superposition of motor-unit action potentials at the skin. It is not a direct geometric measurement of muscle contraction, so joint angles and fingertip force normally require a learned mapping.

Axonnic distributes the sensors across the forearm rather than concentrating them on one muscle group, with the stated aim of reducing overfitting to a small set of dominant muscles. This does not eliminate crosstalk, electrode-to-skin impedance changes, sweating, motion artifacts or anatomical variation between users.

2. Training and deployment use different inputs

The system separates label generation from runtime inference. During training, EMG and hand tracking are synchronized so the tracker can supply joint-angle supervision. At inference, the tracker is removed and the decoder runs from EMG alone. This can reduce visual occlusion during data collection and turn muscle activity into an additional conditioning signal for robotic behavior.

At the interface level, the outputs can be represented as a joint-angle vector θ̂∈R²⁰ and a fingertip-force estimate F̂. An ideal objective could combine angular error, force error and regularization, but the public material does not disclose whether Axonnic uses a joint loss, a shared backbone or staged training. The reported 6.9° angular result therefore cannot be treated as evidence that the force branch has comparable accuracy.

3. Fingertip force still requires individual calibration

EMG can reflect muscle activation, co-contraction, wrist posture and motor strategy, so the same muscle activity does not correspond to a unique contact force. Calibration through force-sensor presses is a sensible starting point, but deployment must test whether grip style, tool, speed and material alter the EMG-to-force mapping.

The training labels are not perfect ground truth either. If joint angles come from hand tracking, tracker localization error, occlusion and jitter propagate into the EMG decoder. Force sensors introduce their own range, drift and latency constraints. For useful robot data, the raw EMG, labels, timestamps and calibration metadata must be preserved rather than storing only decoded joint angles.

Engineering Trade-offs and Reported Performance

  • EMG and vision are not simple substitutes. Video directly observes hand geometry, while EMG captures static grips and subtle effort that may be nearly invisible on camera. Combining them is useful for training, but an EMG-only inference path transfers the accuracy burden to signal quality and individual calibration.
  • 64 channels improve muscle coverage but increase hardware complexity. More channels generally require more analog front-end capacity, sampling channels, wiring and data bandwidth, along with greater assembly, cleaning and electrode-positioning costs. A wired PCB can improve power and timing stability, but cable drag, freedom of movement and long-session wearability still require measurement.
  • Re-wearing is the critical generalization variable. Removing and replacing the sleeve changes electrode locations and skin contact. Axonnic tested separate wearings and reported one 215-second recording made four days after training without recalibration, with a 6.9° aggregate error across 20 joints. That is encouraging, but one recording cannot cover electrode shifts, fatigue, sweating or changes in task distribution.
  • The Apple Vision Pro comparison is not a comprehensive superiority claim. The systems measure different physical quantities, and results can be affected by error definitions, tracking ground truth, joint count, task type and calibration procedure. The material provides no participant count, error distribution, confidence interval, force error, inference latency, power consumption or downstream robotic success rate.
  • End-to-end robotic value has not yet been demonstrated. The public material shows EMG control of a robotic hand and data collection across several manual tasks, but it does not quantify how muscle and force labels improve policy success, sample efficiency or assembly quality. The present system is more accurately described as an EMG demonstration-interface prototype for robot data collection than a validated general-purpose robot controller.

Developer Adoption and Deployment Guidance

  1. Build reproducible wired acquisition hardware first. A custom PCB should prioritize low-noise analog front ends, anti-alias filtering, shielding and grounding, and a shared sampling clock. Mains-frequency rejection should be selected for the actual site spectrum, while motion artifacts and contact-impedance quality should be logged separately.
  2. Treat re-wearing as a formal evaluation condition. Every session should record electrode placement, skin-preparation state, action sequence and calibration version. Results should be reported separately for same-session, cross-session, multi-day and cross-user conditions rather than through a single demonstration.
  3. Synchronize EMG, hand tracking and force sensing rigorously. A shared trigger or hardware clock prevents sensor delay from being misattributed to the model. The data pipeline should also preserve raw signals instead of retaining only decoded joint angles.
  4. Evaluate angle, force and latency together. In addition to mean angular error, report per-joint distributions, force-range error, end-to-end latency, corrupted-frame rates and continuous-running stability. Both no-recalibration and lightweight-recalibration protocols are needed.
  5. Verify downstream robotic gains. Video- or motion-capture-based demonstrations should serve as the baseline. Compare policy success, sample requirements, cross-task transfer and total collection cost to determine whether EMG and force labels are useful modalities rather than expensive redundant inputs.
  6. Keep human review in factory pilots. Abnormal effort, tool contact and material changes in sewing or similar tasks cannot be judged reliably by a single EMG model. Production use therefore needs quality gates, abnormal-operation fallback paths and operator confirmation.

Source: Hacker News AI Feed (2026-11-16)

— END —

Tags: #Axonnic #Surface EMG #Robot Learning #Human-Machine Interface #Dexterous Manipulation

Editorial & Fact-Checking Note: This article is compiled from primary research, official release documentation, and source papers by the ZICQ Newsroom pipeline with automated entity verification and human editorial review. If you notice any technical inaccuracy, please submit a correction via our corrections policy or email our editorial desk directly.

What's your take on this story?
🚀 Explore & Submit Works

Community Comments

Loading live comments and annotations…