And I'd argue that "when" is very, very far away. Unless you remove human drivers, and significantly change the infrastructure for self-driving cars, but at some point you are not so far from (existing) self-conducted trains, then.
Look, people are horrible to one another. They bring their brutalist egos into their cars and drive at the whim of it. Very few people drive the safe distance away from cars, drive safe distances, let alone reacting to perceived slights by other drivers in defensive ways. No, instead they hug bumpers, cut one another off, drive so close a simple pat on the breaks would cause an accident, and so on.
I just wanted to point out that "self driving cars", in isolation without any change to the infrastructure they operate in, are, by my intuition, a pipe dream.
I think we are aligned here.
With what you know about current AI, how it's "trained" instead of "programmed", at the result of something that can at best be analyzed to a certain degree: From the bottom of your heart, would you really, really trust that AI blackbox component to navigate you through real-world traffic situations at dangerous speeds?
20 years is incredibly pessimistic to me. Heck, we might even have GAI by then.
Some problems are really hard, some are essentially impossible. For example, the halting problem is impossible to solve. All its corollaries are, too: You cannot for arbitrary cases tell whether a C program adds two simple numbers together, period.
Voice Assistants are an example for "problems so hard, they are essentially impossible": A human housemate has no problem understanding what I mean when I say "it went yellow, can you order this time?", a voice assistant likely never will. If it can understand that one, it cannot understand another way to phrase it that I can come up immediately. It might get good enough to unambiguously understand "order the water filters for the fridge" every time, but I can convey the same thing to a human in infinitely many more ways.
In driving, unfortunately, you don't get to choose which way a situation is presented, the way you can with crafting a sentence to a Voice Assistant. The next time you drive through a city, look out for any even remotely non-standard situation. Now imagine how a computer should be supposed to react without the arsenal of context clues that your lifetime of human experience provides to you, and that tells you to "just swerve left after slowing down for a few feet" with barely a conscious thought.