Myo Discovery

Every commercial prosthetic hand switches between a handful of pre-programmed grips. None of them respond continuously and proportionally to muscle intent the way a biological hand does.

This started as my undergraduate dissertation in Automatic Control and Systems Engineering at Sheffield in 2015-16, and the core question (can this be done continuously rather than as discrete gestures) never stopped being interesting.

The original iOS app captured EMG via a Myo armband, but hit a hard platform limit: the SDK only exposed processed gestures, not raw signal, which forced a pivot to a Windows tool for raw 8-channel EMG capture, feeding a MATLAB pipeline (bandpass filtering, mean-absolute-value feature extraction, k-NN classification across six gestures). The current system is a Python pipeline synchronising EMG with MediaPipe hand-landmark tracking for ground-truth labelling, built against an explicit five-level movement taxonomy (single-finger, combinations, wrist coordination, transitions between movements, and naturalistic tasks), with transitions being the project's own original contribution.

• Catching an inflated, overly optimistic result in my own early work: accuracy looked near-perfect at first, but testing it properly against data the model hadn't seen dropped that number sharply, and testing across different people entirely, it collapsed completely. Figuring out why mattered more than the number itself.
• Building in real anatomical knowledge, mapping each of the 8 sensor channels to a specific forearm muscle, rather than treating the signal as a meaningless set of numbers.
• Checking the work against what's already been tried rather than in isolation: including the closest prior project at Sheffield itself, and where current published results in this field actually sit, so the project's own contribution is a genuine gap rather than something already solved.

The most important result in this whole project was a negative one: the near-perfect within-session number that fell apart under real cross-validation. Reporting that honestly, and figuring out why, mattered more than any accuracy figure would have.

Knowledge base and synchronised EMG/MediaPipe capture are built. Next: synergy decomposition, continuous EMG-to-joint-angle regression, cross-session personalisation, and eventual hardware interfacing.