emg2tendon: From Muscle Signals to Tendon Control for Dexterous Hands

Created in June 15, 2025

2025   ·   robotics   neural-interfaces   dexterous-manipulation   deep-learning   ·   research

The Control Problem for Tendon-Driven Hands

Tendon-driven robotic hands are far more dexterous than joint-actuated designs — but that dexterity is hard to command. Motion-capture data doesn’t map cleanly onto tendon controls, and vision-based hand tracking breaks down under occlusion and self-contact. Surface electromyography (sEMG) — reading the electrical activity of forearm muscles from a wrist-worn band — is a robust, inexpensive alternative that’s available even when the hand is out of view. The catch, and the core of this project: mapping raw sEMG to tendon control forces remains an open problem.

emg2tendon is my independent research effort (RSS 2025) attacking exactly that mapping.

The Approach

The pipeline bridges neuromuscular signals to mechanical control:

  1. Record sEMG from subjects performing a wide range of hand gestures.
  2. Use the MyoSuite MyoHand musculoskeletal model to derive the corresponding tendon control signals — giving supervised targets that are biomechanically grounded.
  3. Train regression models to predict tendon forces directly from the EMG.
  4. Introduce a novel diffusion-based regression approach for the prediction itself.

The Dataset

A central contribution is the EMG-to-Tendon Control dataset, which extends the prior emg2pose data with tendon-control annotations:

  • 193 subjects
  • 370 hours of recordings
  • 29 diverse gesture categories — thumbs up/down, grasping, pinching, and free-form movements across 29 gesture stages.

To make labeling at this scale tractable, a C++ QForce tendon implementation delivers a ~20× speedup over the Python pipeline, processing 1,700 hours of motion-capture data in 24 hours.

Why It Matters

This is the neural-interface half of my Micropilot work: closing the loop from human muscle intent → tendon forces → musculoskeletal actuation for anthropomorphic hands. Getting a robust sEMG-to-tendon map is what makes it possible to teleoperate — and eventually learn policies for — dexterous hands from a cheap, wearable signal rather than a motion-capture rig.

Project page: emg2tendon.github.io.