Answers are grounded in this paper and cite their sources.
Gesture recognition from electromyography (EMG) is used to control prostheses and rehabilitation devices, but real-time classifiers must balance accuracy, latency, cost and computational load. This paper presents a tiny machine learning (TinyML) pipeline that classifies hand and arm gestures from envelope EMG on a low-cost microcontroller. Four-channel envelope EMG was recorded with a six-lead myoelectric module and an Arduino UNO from 10 healthy participants, who performed six gestures for 4 s each in two sessions. A joint accuracy-complexity search selected the window length, overlap and four time-domain features used to train an artificial neural network (ANN). On a 60/20/20 window-level training/validation/test split, the test accuracy was 98.35%. Five-fold cross-validation (97.37 ± 1.05%) indicated stable performance across data partitions, and leave-one-subject-out (LOSO) evaluation, the only subject-independent test, gave 97.22 ± 1.33%. Compared with a support vector machine (SVM), k-nearest neighbours (KNN), a decision tree and a random forest, the ANN had the lowest golden distance (GD = 0.0325). Deployed on a Raspberry Pi Pico in TensorFlow Lite format, the model uses 9.1% of flash memory and an estimated 10.6% of static random-access memory (SRAM), with a 23 ms inference time, a 254 ms total on-device cycle and a 12.9% active processing duty cycle. In a representative real-time test, with a PC relaying the EMG data to the Pico, per-gesture accuracy over 500 consecutive predictions was 95.8–99.6%. These results show that real-time envelope-EMG gesture recognition is feasible on a resource-constrained microcontroller; with validation on larger and more diverse populations, the approach may inform human-machine interfaces for prosthetics, rehabilitation and assistive devices.
No takes yet. Share an insight, caveat, or question.
KHAN et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: