I think a lot of people in this thread are underestimating the possibility of safety drivers going remote and how that could be a solution to these edge cases.

When making decisions or coming across strange situations the AI has a confidence level, and when that falls below a threshold it can notify a remote person to potentially handle the situation. Right now Waymo has problems merging on a freeway with asshole drivers... but what's the big deal? It's maybe 1 minute out of a 30 minute drive.

If you only need a human operator to take over the vehicle 2 minutes out of every hour of autonomous travel, then you could probably get by with 10 remote operators monitoring a fleet of 200 vehicles. That could translate into way better profit margins than any current ride-share service.