The Edge Cases That Still Trip Up Autonomous Vehicles

A handful of scenario types keep showing up in the industry's public failure reports, again and again, across companies that otherwise have very little in common technologically.

Construction-zone detours are one of the most persistent categories, especially when the signage is temporary, hand-lettered or contradicts what a map or a permanent sign says elsewhere on the same block. Unprotected left turns across oncoming traffic are another: judging when a gap is wide enough to cross safely, with no signal telling either side what to do, is one of the harder real-time decisions in ordinary driving, let alone automated driving. A third recurring category involves emergency vehicles and first responders, where a system has to recognize lights, sirens and unusual vehicle behaviour fast enough to yield correctly. What ties the three together is ambiguity: each one is a situation where the normal rules of the road don't fully specify what to do, and a human driver would ordinarily be filling the gap with judgment rather than a fixed procedure. Timing matters more than recognition.

A fourth category worth separating out from the rest involves objects that emerge from occlusion rather than ones that were simply hard to classify once visible: a pedestrian stepping out from between two parked delivery vans, a cyclist appearing from a driveway hidden behind a hedge. The perception problem there isn't recognizing what the object is once it's visible, cameras and lidar are generally good at that; it's the narrow window between when the object becomes visible and when it's already close enough that a late reaction has real consequences. That's less a sensing failure than a prediction and reaction-time problem, and it shows up in ordinary human driving too, which is part of why defensive-driving guidance spends so much time on exactly this kind of occluded approach.

It's also worth distinguishing failures where a system didn't see something from failures where it saw the situation correctly but misjudged what to do about it. An unprotected left turn is rarely a perception failure in the narrow sense, the oncoming vehicle is plainly visible, the gap is plainly measurable; it's a judgment call about acceptable risk that even skilled human drivers make imperfectly. Lumping every edge case into a single sensor-failure narrative obscures that a meaningful share of these incidents are prediction and decision problems sitting downstream of perfectly good sensor data, which points toward a different kind of fix than simply adding another camera or radar unit.

What makes this pattern notable is that it isn't specific to one company's technology stack. These same categories recur across multiple operators' public disengagement and incident disclosures, which suggests the difficulty sits in the scenarios themselves rather than in any single approach to solving them. A pattern that shows up across several unrelated technology stacks is a pattern in the driving task itself, not a coincidence traceable to any one company's training data.

That's also close to the point of disengagement-reporting requirements in the first place. Regulators didn't design them purely as a transparency exercise. They exist partly to surface exactly this kind of cross-industry pattern, rather than leaving each company to grade its own homework and report only what it chooses to. The disengagement reports compiled from those disclosures are where this kind of pattern actually becomes visible instead of anecdotal.