Self-driving cars that stop when confused may be trading one risk for another.
A paper posted to arXiv examines publicly documented incidents where autonomous vehicle fallback behaviors — the tendency to slow or halt when the system hits uncertainty — ended up obstructing traffic, blocking emergency responders, and stranding passengers. The researchers categorize these failures across three layers of the standard AV stack: perception, planning, and control. Their conclusion is that the industry's dominant safety logic, built around collision avoidance and what engineers call minimal risk conditions, does not account for how human-governed roads actually work.
The gap matters because city deployments are no longer hypothetical. When a robotaxi freezes in an intersection or pulls over at an inopportune moment, it doesn't just inconvenience drivers — it can delay an ambulance or leave a passenger with accessibility needs stranded without recourse. The paper argues that AVs need to interpret human authority signals, respond to multimodal instructions, and adapt to social traffic norms, not just physical obstacles.
The proposed fixes — remote teleoperation, language-grounded planning, accessibility-aware routing — are all active research areas, so this reads more as a gap analysis than a solved problem. What the paper does usefully is name a blind spot: the AV industry has spent years optimizing for not crashing, and rather less time on what happens after the car decides to stop.