It's easy to make polarized judgements, like "AI driving can only follow lanes or drive on wide broad sunny streets". I recently watched a lecture where a guy from Waymo talked and demoed some situations, check out the 3 seconds or so in the timestamp in this video which made me re-evaluate that pessimism a bit:
Sure it's still nowhere near the general understanding necessary for completely autonomous driving but still much more impressive than simple lane-following.
That clip is far from convincing. Sure, the video showed an AI making the correct decision in a single complex circumstance. But that is assuming you have a correctly-modelled situation with accurate sensors. Even if you could guarantee that the AI could make the correct decision every single time, millions of times per year, you have the problem of potentially inaccurate and/or failing sensors which give the AI an incomplete model.
This is why airplane auto-pilots disengage when they detect failing sensors, giving humans enough time to take over. Unfortunately in the case of cars, there is not enough time to hand over to a human.
It isn't hard to have redundant sensors so that you can drive for a bit on less than all of them. If there is a failure you drive to a mechanic and get the senor fixed before there is a second sensor failure.
The odds of a dual sensor failure is low enough to still be substantially better than humans. Assuming of course that the dual sensor failure is dual unrelated failures. That means care in the wiring of power systems, fail safe modes, over design such that failure can't happen, and other engineering tricks.
https://youtu.be/Q0nGo2-y0xY?t=865
Sure it's still nowhere near the general understanding necessary for completely autonomous driving but still much more impressive than simple lane-following.