Why do we even need such accuracy? Humans are extremely bad at estimating distance and nevertheless are quite ok drivers. ML needs to make sure that the car is not bumping at things or people, which is a very different task.

Because the failure modes are different. Humans with basically 100% accuracy both correctly recognize depth and put it into the right bucket of near/far but are woefully bad at estimating the numeric distance. So you’re right that you don’t need to do better than extremely coarse estimates like close, a few car lengths, in the distance, etc..

If computer vision fails at all to recognize depth (semi truck with clouds painted on it isn’t recognized) or misclassifies whether something is near or far that’s much more dangerous.

Yes, so "can/cannot bump into" is a classification tasks which needs quite some accuracy.

A human on the other hand can recognize another human with near perfect accuracy.