Locomotion-State Classifier

I trained a locomotion-state classifier on hip-exoskeleton encoder data across eight trials, reaching roughly 98.7% agreement with hand-labeled ground truth.

The harder part was validating it honestly. Splitting windows randomly across trials lets adjacent strides from the same trial land in both the training and test sets, so a model can look accurate by memorizing a trial instead of generalizing across gaits. I used grouped cross-validation, holding out entire trials rather than individual windows, so every test fold was on a trial the model had never seen. It's a harder bar to clear, and the score that comes out of it means something.