The edge case the patent quietly admits is the one that defines autonomy: people do not move the way the model predicts. US11467580B2, granted to UATC in October 2022, is "Systems and methods for detecting surprise movements of an actor with respect to an autonomous vehicle."

Most AV prediction works by forecasting what nearby agents will do — the pedestrian will keep walking, the car will hold its lane. Classified under B60W 60/0027 (autonomous operation with behavior prediction) and G05D 1/0221, this patent inverts the focus: it is about noticing when an actor violates the forecast. The cyclist who swerves, the pedestrian who darts back, the car that brakes unexpectedly.

“Systems and methods for detecting a surprise or unexpected movement of an actor with respect to an autonomous vehicle are provided.”— U.S. Patent No. 11,467,580 source

The mechanism in claim 1 is built around two compute cycles and a precomputed safety zone. In a first cycle, the system obtains motion-forecast data for an actor relative to the vehicle and from it determines "failsafe region data representing an unexpected path or an unexpected area where a likelihood of the actor following the unexpected path or entering the unexpected area is below a threshold." That is the key inversion stated as engineering: rather than only predicting where the actor will go, the system explicitly computes the regions the actor is predicted not to go — the low-probability tail — and labels them as failsafe zones to watch. In a later cycle, it takes fresh sensor data, checks whether the actor has in fact entered one of those unexpected areas, and if so "initiat[es] a second processing flow" and computes "a deviation for controlling a movement of the autonomous vehicle." The surprise is detected as a boundary crossing into a pre-marked zone.

How the failsafe region is drawn is specified. Claim 7 ties it to geometry: the system identifies "an intersection of an actor trajectory… with a trajectory of the autonomous vehicle" — the unexpected area that matters is the one where the actor's improbable path would cross the car's own. Claim 3 detects the violation by spotting "spatial data points that indicate that the actor has followed the unexpected path." And claim 9 names the trajectory models used to forecast the actor's continued motion once it has surprised the system — "a ballistic trajectory or a quadratic regression trajectory," lightweight motion extrapolations chosen for speed in the reaction path rather than a heavy learned predictor.

Here is why detecting surprise is harder and more valuable than predicting normalcy. A prediction model is optimized for the common case, and the common case is safe almost by definition. The danger lives in the tail — the movements the model assigned low probability — and a system that explicitly watches for prediction failure is watching exactly where the crashes are. The two-cycle structure is what makes the reaction fast: the failsafe region is computed ahead of the surprise, so when the actor crosses into it the system is matching against a precomputed zone rather than re-reasoning from scratch.

That speed emphasis runs through the dependent claims, and it reveals the architecture's real purpose. Claim 4 lets the second cycle "process[] a portion of the second sensor data that corresponds with the unexpected area without processing all of the second sensor data" — look only where the surprise was flagged. Claim 10 says the second processing flow "omits processing by at least a portion" of the systems used in the first — a deliberately shortened, lower-latency reaction path. Claim 6 precomputes "a preliminary deviation" before the surprise is even confirmed, so a response is half-ready the instant the actor crosses the line. Claim 5 even allows the system to retask a sensor to "collect additional data that corresponds with the unexpected path." Read together, this is a fast reflex arc bolted alongside the normal perception-prediction-planning loop: spend less compute, look in fewer places, and have part of the maneuver staged in advance, because the whole value of the feature is shaving latency off the response to the unmodeled human.

Claims 2 and 14 show what the deviation actually is in many cases: not necessarily a fresh emergency plan computed under time pressure, but a previously prepared "motion plan that is determined during the first compute cycle" that replaces the current plan when the surprise fires. The car keeps a contingency motion plan ready and swaps to it — a documented form of always having an out.

The honest framing is that surprise detection is an admission of prediction's limits. If forecasting were perfect, you would not need a separate mechanism to catch its failures. Patenting the surprise detector concedes that the prediction model will be confidently wrong about actors, and builds a safety reflex for that moment. It also inherits prediction's weakness in a subtler way: the failsafe regions are only as good as the forecast that drew them, so a sufficiently strange actor could surprise the surprise detector. The patent narrows the risk; it does not eliminate it.

This is the kind of unglamorous safety logic that does not appear in autonomy marketing, which leans on smooth highway footage. Real-world disengagements cluster around exactly these surprise events — the unmodeled human — and the IP that addresses them is a better window into AV maturity than any mileage figure. That UATC (the Uber autonomy entity whose assets moved to Aurora) filed it is itself a tell about where serious AV engineering effort goes: not into the demo, but into the reflex for when the demo's assumptions break.

For readers auditing self-driving claims, the surprise-movement question is the sharp one: not "how well does your system predict normal behavior," but "what does it do when behavior is abnormal." This patent is one company's documented answer — precompute the unexpected zones, watch only them, keep a contingency plan staged, and react on a deliberately shortened compute path — and the documentation matters.