Not only that, but if I were in charge of picking routes for autonomous test vehicles to drive, I'd pick scenarios that the cars have trouble with. Except for

Not only that, but if I were in charge of picking routes for autonomous test vehicles to drive, I'd pick scenarios that the cars have trouble with. Except for a small amount of regression testing, not much point in sending it out to deal with scenarios I know it can already handle. I'd want to gather data on whether my software improvements have given it the ability to handle scenarios it previously failed on. Point being, it's probably not a random sample of possible routes. It's probably skewed. So the statistics probably aren't very meaningful.