Toward Plug and Play Myoelectric Control via One-Shot Latent Representation Alignment

Myoelectric control has emerged as a common tool for human-machine interaction with applications in prostheses, rehabilitation, and mixed reality. Despite the growing research and promising results in controlled laboratory settings, there are challenges to its widespread adoption. During activities of daily living, factors such as limb position, electrode shift, cross-day variability, and individual differences reduce the reliability of controllers. Moreover, most existing work to address these issues is validated on individuals without limb differences, limiting clinical generalizability. We posit that even in the presence of these confounding factors, the intention of performing a specific gesture remains similar. Hence, a common underlying dynamics of muscle activity could be leveraged to create a robust myoelectric controller. Therefore, we propose a one-shot learning framework based on Multi-set Canonical Correlation Analysis to align the latent representation of the surface electromyography signals to achieve reliable myoelectric control across limb positions, days, and individuals with only minimal calibration data from a new condition (i.e., limb position, day, or individual). Importantly, we show that our framework generalizes from individuals without limb differences to an individual with a congenital limb difference despite different muscular physiology. Therefore, our framework can eliminate the need for retraining and data-hungry models, promoting plug-and-play myoelectric control robust to variations in limb position, day, and individual.