EMGFlow: Robust and Efficient Surface Electromyography Synthesis via Flow Matching
New AI model creates synthetic muscle signals for prosthetics, outperforming GANs and diffusion models in speed and accuracy.
A research team led by Boxuan Jiang, Chenyun Dai, and Can Han has introduced EMGFlow, a novel AI framework that generates synthetic surface electromyography (sEMG) data using Flow Matching technology. This represents the first application of continuous-time generative modeling specifically for muscle signal synthesis, addressing a critical bottleneck in gesture recognition systems where real-world data is scarce and lacks subject diversity. The model was rigorously tested across three benchmark sEMG datasets using a unified evaluation protocol that measured feature fidelity, distributional geometry, and downstream utility.
EMGFlow consistently outperformed conventional data augmentation methods and Generative Adversarial Network (GAN) baselines, while demonstrating stronger standalone utility than diffusion models under the train-on-synthetic, test-on-real (TSTR) protocol. By optimizing generation dynamics through advanced numerical solvers and targeted time sampling, the framework achieves superior quality-efficiency trade-offs compared to existing approaches. The researchers have made their code publicly available, providing a practical tool for developers working on myoelectric control systems for prosthetics and human-computer interfaces.
- First application of Flow Matching for sEMG synthesis, outperforming GANs and diffusion models in TSTR protocols
- Achieves improved quality-efficiency trade-offs through optimized numerical solvers and time sampling strategies
- Addresses critical data scarcity in gesture recognition for prosthetic controls and human-computer interfaces
Why It Matters
Enables more robust training of prosthetic control systems by generating diverse, realistic muscle signal data where real-world collection is limited.