Real world evidence would suggest that human-level cognition is not enough to safely drive a car

And yet, as of today self-driving cars aren't nearly as good as a human even after 2-3 beers.

It's not that black and white. The unnerving thing is that self-driving cars have completely different failure modes than drunk humans: instead of falling asleep and drifting into the next lane, they do (to humans) absurd things like mistake the side of a tractor-trailer for empty space and merge into it.

I firmly believe that the literature produced which shows self-driving cars are safe or even safer, and there is scads of it now, is mainly down to the program's ability to make sure a safe following distance is observed leading to greater reaction time for braking. I know a lot of driving tests have been done in simulators which show the AI can be safer but I agree with your statement.

I also have very little confidence in the AI's ability to stop at road construction, make its way over temporary embankments and diversions placed for construction, navigating in inclement weather or adverse traction situations like a muddy drive, etc. I suppose the real goal here is for Wal-Mart and Amazon to move freight on the Interstates to begin with but I don't even see that happening any time soon. Your average Wal-Mart loading bay and parking areas are likely much more than any current AI can handle anyhow. How does the AI know the person in the mart cart isn't going to just jut out in front at the last second? Can't exactly make eye contact.

Hmmm.. Not accurate? At all?

Personally, I don't think so. If anything we might need less machine learning.

It's mostly a problem of sensor fusion. Once your self driving car has a good view and solid understanding of the surrounding environment, the actual driving algorithm is reasonable simple. All you really need is really solid software engineering to implement it with minimal bugs.

The problem is getting that understanding. You can't have objects pop in and out of existence. You can't have computer vision misclassifying objects or just not seeing them.

It might just require better sensors, or more types of sensors. Or we might need to work on interpreting the data and understanding the actual situation.

Most likely it's a combination of both.

I believe so. There are way too many variables when it comes to driving. I honestly don't think it's possible to solve the problem without huge advances in general purpose AI that can do its own goal setting ... and then I just imagine Marvin from Hitchhiker's Guide...

"...one of the most advanced robots in the galaxy that can do billions of computations a second and you ask me to do the most mundane tasks... go open the door Marvin. Go pilot the ship Marvin. Go fix the reactor Marvin ... Go drive my Tesla Marvin ... I think you ought to know I'm feeling very depressed..."

Not on a static road without people, animals, weather, geological events, etc.

So replace roads with tubes containing a fixed track that prevents animals, weather, and people from interfering. Then add tube outlets to areas that are frequented by most travelers. You could even put these tubes underground to save space, and instead of large individual cars you could optimize the vehicles for many passengers in a small amount of space.