Question

Design the data + labeling pipeline for a self-driving perception model. You collect petabytes of sensor logs from a fleet, but only a tiny fraction is worth labeling (most frames are boring highway driving), human labeling is expensive and slow, and rare dangerous scenarios (a child running into the road) are exactly what you most need but rarely captured. Design the pipeline from raw fleet data to a labeled training set, focusing on how you decide what to label and how you keep getting better at finding the rare, valuable cases.