You cannot test-drive your way to proof of safety.
Automated driving systems fail on the rare mile, not the average one. Those miles are too rare to observe enough of.
OutRes (Outlier Resilience LLC) is an early-stage applied research venture working on that problem.
01The arithmetic doesn't work
Human drivers average 1.09 fatalities per 100 million miles. RAND calculated what proving an automated vehicle beats that would take.
“It is not possible to test-drive autonomous vehicles to demonstrate their safety to any plausible standard, even if we assume perfect performance.”
Kalra & Paddock, Driving to Safety, RAND, 2016
02The standards name the gap, not the answer
ISO 26262 covers hazards from things breaking. But a system can cause harm with every component working as designed. The design was insufficient for the situation.
ISO 21448 (SOTIF) covers that: risk from functional insufficiencies, no failure required. It asks you to shrink the set of unknown-unsafe scenarios. It does not tell you how to know how much you have eliminated.
03The tail is not exotic. It's Tuesday in February.
The conditions that degrade a perception stack are mundane, frequent and regional. A test program assembled somewhere sunny will under-weight them.
- Degraded optics
- Low sun, wet glare, spray, a salted lens. The image still arrives, degraded.
- Vanishing road structure
- Markings under snow or wear. Models that lean on paint inherit a dependency the road doesn't keep.
- Shifted statistics
- A model fitted to clear conditions meets one it wasn't trained on. The failure is quiet.
- Compounding conditions
- Dusk, precipitation and a worn surface at once. One-factor-at-a-time matrices miss the combinations.
04Finding out late is the expensive way
Early, a weakness is a design decision. In hardware-in-the-loop or on-road validation it is a schedule event. So the aim is to move the evidence earlier, onto conditions you have not sampled.
05Research
Claims here should arrive as papers with methods attached. Work in preparation will be posted here.
06Compare notes
If you run validation, simulation or safety for an automated driving program, we'd like to hear where your process runs out of evidence. We're not selling anything yet.