EMG-based Continuous Trajectory Prediction using CNN-LSTM Networks

Current camera-based hand tracking is limited by occlusion and environmental factors, hindering practical applications in robotics and digital interaction. Although electromyography-based tracking offers a robust alternative, a fine movement regression, such as pen trajectory prediction, remains challenging. This study addresses the problem of predicting 2 dimensional axes’ velocities of a pen using 8-channel forearm electromyography signals. We propose a novel deep learning approach based on the decomposed electromyography methodology, which transforms complex raw signals into latent and independent muscle activation patterns for effective movement command learning. Furthermore, we integrate a signal reconstruction mechanism during the training process. By minimizing the reconstruction loss, the model is jointly optimized for both velocity decoding accuracy and physiological fidelity of the signal representation. This approach allows for the experimental validation of a model that utilizes the muscle signals alone to predict the continuous pen input, demonstrating the feasibility and potential for advancing camera-free digital handwriting and human-computer interaction.