I was driving home a few months ago in the rain on Division in SF. Cars were stopped, and I immediately couldn't figure out why. When I passed the stopped vehicles I got my answer: there was seemingly a wall of water coming down from the freeway overhead and an autonomous car was stopped as it apparently didn't know how to handle the situation.
Would need incredibly low latency for that to be safe, I think. FPS games use a pile of tricks to send minimal data over the network which wouldn't be available to an autonomous car: we don't have access to the game engine, after all...
You wouldn't automatically need low latency. In a situation like the above one the car could pull over and wait for the control center to tell it whether the drive on through the water or not.
That would require the car to have the ability to override its tendency to not drive thru physical obstacles. Remote control to say "go into that lane", "turn left" is one thing, but remote control to drive thru what it perceives as a physical object is different.
It would need to be very low latency if the cars can not take sufficient action fast enough to be safe. But if the cars get to a point where they are safe enough but where that safety comes at the expense of being overly cautious in some instances (refusing to drive through a sheet of water and deciding to stop) that may be acceptable. E.g. in the given example, the car only needs input to say pretty much "safe to proceed at low speed until past obstacle, then reassess".
Hopefully not, at least not yet. That doesn't sound nearly as safe, or as accountable— you just wouldn't have the same situational awareness looking at sensor feeds vs. looking out the actual windows.
