Muscle control of an extra robotic digit

Abstract

Controlling an extra robotic finger requires the brain to adapt existing motor signals. While most current strategies exploit physical movement, there is growing interest in harnessing muscle activity directly via surface electromyography (EMG) as a more seamless interface. We systematically compared muscle- (EMG) and movement-based (force sensor) control of a Third Thumb. Using identical instructions and a counterbalanced within-participants design, we assessed initial skill, learning, and cognitive load across a variety of tasks, enabling a blinded comparison across control modalities. Both control modalities afforded successful Third Thumb control and learning, although force control consistently delivered better performance. Despite execution differences, learning rates and cognitive loads were comparable, with a similar evoked sense of agency. Signal analyses showed performance was predicted by real-time force sensor parameters but not by EMG, reflecting distinct control dynamics. Nonetheless, EMG training led to greater skill transfer to force control, suggesting it may better support generalisable learning. These findings challenge the assumption that proximity to neural signals ensures better control. Although EMG underperformed in execution, it showed unique advantages, including enhanced generalisation and access to richer signals, highlighting the need for improved real-time decoding to fully exploit its potential.