sEMG-Based Wearable HMI System For Real-Time Robotic Arm Control With Edge AI

Abstract

Gesture and motion recognition based on surface electromyographic (sEMG) signals brings intelligent and natural interaction experience for human-machine interaction (HMI). Deep learning (DL) based recognition methods are currently widely used. However, performing deep learning in real-time on resource-constrained edge portable devices is still in its infancy. Therefore, there is a great demand for efficient, accurate, cost-effective solutions for edge devices. To address this challenge, we built a low-power and wearable HMI system with edge AI. Specifically, we proposed a multi-channel sEMG signal-based gesture and action recognition algorithm, developed a portable 8-channel sEMG acquisition circuit, and designed a dedicated algorithm acceleration processor to implement the proposed algorithm. The data of 10 categories in 6 healthy subjects were tested with an average recognition accuracy of 95.73%. In addition, we built a robotic arm online test system for testing, and the experimental results showed an average success rate of 92.5%, a single-frame signal processing and recognition time of 44.36 s, and an overall system power consumption of 0.493W. This system meets the needs of real-time applications and can be integrated or extended to other application scenarios.