1D CNN enables real-time muscle-controlled robotic arm with 90%+ accuracy

A team led by Edgar Manacorda has demonstrated a real-time surface electromyography (sEMG) interface for teleoperating assistive robotic arms, leveraging a 1D convolutional neural network (CNN) to classify muscle signals. The system captures four-channel EMG data, preprocesses it, and segments it into sliding windows for classification. Two approaches were tested: a two-stage classifier (rest vs. movement followed by gesture recognition) and a single classifier handling both rest and five distinct gestures. The CNN-based approach consistently achieved over 90% test accuracy on experimentally acquired signals, with a stable real-time latency of approximately 0.32 seconds—well within the windowing strategy limits.

In both simulated environments and on a physical robotic platform, the system translated user muscle activity into smooth, coherent robot movements. The research highlights the feasibility of deep learning-powered sEMG control as a non-invasive, intuitive interface for individuals with upper-limb motor impairments. By integrating signal processing, classification, and control into a unified real-time framework, the team demonstrated that discrete gestures can reliably guide robotic arms for daily assistance tasks such as object manipulation.

The paper, published on arXiv (ID: 2607.16310), notes that future work could combine sEMG with additional sensing modalities to improve robustness and usability. This advancement moves closer to practical, affordable assistive robotics that respond to muscle signals without requiring surgery or extensive calibration.