[malozite]

I think the data advantage is limited.

I used to work in high-frequency equity trading. We would run simulations against tick-by-tick data from real markets. The simulations were of very questionable value. As soon as you trade once in your simulation, from that point onwards the whole run is contaminated. Your trade, if it had actually happened, would have been seen by other participants, who would have changed their strategy, often in complex and unpredictable ways, and the timelines would have diverged rapidly and irreversibly from that point.

Data recorded from driving situations is useful for one thing - testing out split-second decision making - stop or accelerate? Steer left or right? This is a limited use case. Most situations don't have critical decisions like this. The ones that do, it's often hard to judge what would have been the right answer. Even accidents - where the post facto desired outcome is "don't have an accident", unfold over several seconds, involving several stages of actors responding to each others' actions. No one can be sure how the driver should have approached the very first interaction - the only one for which the data is meaningful.

Actual road testing, and meaningful simulation which isn't just replaying what actually happened is much more useful.