It is 2017, and autonomous vehicles are on public roads in Pittsburgh and Phoenix with safety drivers whose attention is the only backstop. Perceive and predict everything around the car — vehicles, cyclists, the pedestrian stepping off the curb mid-block — in rain, at night, against low sun, with a perception stack whose benchmark numbers were earned in daylight. The technical crux is the tail of the distribution: the jaywalker with a bicycle at 10 p.m. that appears in no training set, the distribution shift between the mapped test city and everywhere else. Ninety-nine percent accuracy is a fatality every few weeks at fleet scale. Within a year, a pedestrian in Tempe, Arizona will be killed by exactly this failure. Build the stack as if you know that.
Nowak's active learning and sparse recovery work bears on exactly the kind of question a perception system faces at the distributional edge — how to efficiently characterize rare, informative examples like the jaywalker with a bicycle — giving him a real if narrow methodological connection. He did not build automotive systems, so this is a signal-processing-adjacent theoretical fit rather than deployed experience. Nowak's recognized contribution — active learning; sparse recovery; signals meet machine learning — sits adjacent to this problem's core demands rather than inside them, which is why the fit reads as partial credit: real intellectual kinship, no applied instrument for the task at hand. Placed in the roster's stat-learning tranche under ml, the surrounding record gives no further grounds to move this score: the specific instruments this problem asks for are outside the body of work that made Nowak notable.
He has taught perception and robotics case studies at Berkeley, sat in rooms with people who ship autonomous vehicle stacks, and can explain distribution shift on a whiteboard with real fluency. Then Urtasun walks in having actually built and iterated the perception stack across a real fleet in exactly this window, Malik supplies the segmentation theory underneath it, Girshick supplies the detection architecture inside it, and Hamilton supplies the safety-critical software discipline the whole system needs to not kill anyone. The professor's contribution is a very good lecture slide. Zero.
Battle #29 · 8/9/2026, 6:20:08 PM · this result is deterministic: the same two personas on this problem always resolve the same way.