GNN model achieves 99% real-time gesture recognition from sEMG signals
New graph neural network reads muscle signals with 99% accuracy in 48ms.
A team of researchers from the University of Miami has introduced a novel approach to real-time hand gesture recognition using surface electromyography (sEMG) signals. Led by Pragatheeswaran Vipulanandan, Kamal Premaratne, and Manohur Murthi, their method converts raw sEMG data from an 8-electrode MyoBand into graph networks that capture spatial relationships between muscle activation patterns. A graph neural network (GNN) then classifies gestures from these representations, leveraging the structural information of muscle activity for higher accuracy.
Evaluated on 8 healthy subjects, the model achieved a 99% average classification accuracy, outperforming state-of-the-art techniques. Crucially, the entire pipeline—graph construction plus prediction—averages only 48ms on an Apple M1 Pro CPU, meeting real-time requirements for seamless control of advanced hand prostheses and augmented reality interfaces. The work, published on arXiv, demonstrates that graph-based sEMG representation can significantly boost gesture recognition performance while keeping computational demands low for practical deployment.
- 99% average classification accuracy on 8 subjects using an 8-electrode MyoBand
- Total graph construction and prediction time of 48ms on an M1 Pro CPU, enabling real-time use
- Novel graph representation of muscle activation patterns improves over traditional sEMG methods
Why It Matters
Boosts prosthetic control and AR interaction with near-perfect, real-time gesture recognition from muscle signals.