How we turn hours of head-camera footage of people at work into clean, labelled datasets with 3D hands that robot-learning code can load directly.