A robot can theoretically drive better than a human because emotions and boredom don't have to be involved. But we aren't there yet and Teslas are trying to solve the hard mode of pure vision without range finding.

Also, I suspect that the ones we have are set up purely as NNs where everything is determined by the training, which likely means there's some random-ass behaviour for rare edge cases where it "thinks" slamming on the accelerator is as good an option as anything else but since it's a black box no one really understands, there's no way to tell until someone ends up in that position.

The tech still belongs in universities, not on public roads as a commercial product/service. Certainly not by the type of people who would at any point say, "fuck it, good enough, ship it like that", which seems to be most of the tech industry these days.