Read three robotaxi planning patents issued on the same June day and a single edge case keeps surfacing. It isn't perception — these aren't about seeing. It's prediction: how does the car model what every other human on the road is about to do? Zoox, Aurora, and Nissan each filed a different answer to that one question.
Aurora's grant US12649490B2, "Systems and methods for generating behavioral predictions in reaction to autonomous vehicle movement" (CPC B60W 60/0011, G06N 3/08), names the hardest version outright: other drivers react to the robotaxi. The car merges; a human behind it brakes or accelerates in response. A planner that predicts other agents as if they ignore the AV is wrong in exactly the situations that cause crashes. Aurora is patenting prediction that closes that loop.
Zoox's US12649487B2, "Driving surface cost landscape," comes at it from geometry: turn the drivable space into a cost map where some paths are cheap and others — too close to a pedestrian, into oncoming risk — are expensive, then plan the lowest-cost route. The prediction of other agents shows up as cost. Nissan's US12649492B2, "Constraint-based speed profile" (CPC B60W 30/0956, B60W 40/105), works the speed axis: choose how fast to go subject to constraints that other agents impose.
Vision-only versus mapped is a cost bet — but prediction is the bet nobody escapes. Whatever your sensor religion, once you've perceived the scene you still have to guess the future, and the future contains humans who are themselves guessing about you. That recursive prediction is the genuinely unsolved core of urban autonomy, and the fact that three independent teams patented three angles on it the same week is the clearest signal of where the frontier actually is.
What none of the three claims is a measured safety margin. Cost landscapes, behavioral predictions, and speed constraints are methods; how often they correctly anticipate a jaywalker or an aggressive merge is validation data, not claim language. The edge cases the patents quietly admit — reactive agents, ambiguous intent — are precisely the ones where method and reality diverge.
For readers refereeing the robotaxi field, the takeaway is to stop scoring perception demos and start asking about prediction. These three filings agree, across three companies, that modeling other people is the part that's hard — and the part still being invented.