This paper explores non-invasive use of surface electromyographic (sEMG) signals from a human arm for controlling various devices. In recent years, numerous studies have explored sEMG-based gesture recognition for prosthetic and robotic applications. Gesture sets are often limited to a small number of movements, restricting the range of control. Furthermore, the translation of offline classification results to real-time robotic control remains challenging due to latency, signal variability, and computational overhead. These limitations motivate further research into robust, adaptable, and computationally efficient sEMG-based control systems. An 8-sensor Myo Armband device is employed for sEMG signal acquisition. This study involves four participants—comprising an equal number of men and women with diverse ages and body conditions. Each participant performed nine different gestures, repeated 10 times, yielding a comprehensive training dataset. Various machine learning algorithms were applied to filter the raw signals, scale the data and classify the gestures using optimized parameters. During evaluation, the most effective filtering methods and classifiers (with subject-specific tuning) were selected for near-real-time gesture classification and robotic device control. The trained model predicts the performed gesture from the nine available classes and transmits the corresponding command to the robotic system. The chance or baseline accuracy of the system thus translates to 11.1% with the probability of randomly selecting one correct gesture out of nine. The evaluation phase demonstrated that Random Forest, Linear SVM, and Extra Trees were the top three classifiers. The classification model achieved an average accuracy of 92.8%. Despite the promising results, real-time classification for robotic control remains a challenge, necessitating further refinement of gesture segmentation and signal processing techniques.
Valcheva et al. (Thu,) studied this question.