Keep in mind, they don't have to be perfect!
All they have to do is beat we humans, who did not evolve to be optimized for continuous attention to something that is more or less the same as it was for teh last hour (driving interstate).
As an analogy, when I was doing work in Human Factors, read some research where they needed a human to visually monitor test tubes passing by, to determine the results of blood work (I think computer vision systems could easily perform this function now). This was a medical test lab and they had a very difficult time getting the high level of reliability they needed. The positive indication (for the relatively rare disease they were testing for) was an off-color which was easily distinguished, but not a dramatic contrast from the others (let's say it was blue-grey among grey samples). They ended up solving the problem by "seeding" the line with "dummy positives".
(I don't remember the exact times/numbers, but the concept will be clear)
Before seeding, a "positive" would show up randomly, but at an average of once every two hours! This was too long to hold the inspector's attention, and they would lapse into daydreaming. By mixing in "dummy positives" such that positives were presented every 10 minutes on average, the inspector could maintain his vigil at a much higher level and the reliability of detection went up dramatically! I don't believe they told the inspector about the dummies, but he also found his job more rewarding after the trickery!
So I guess the conclusion is, if we want to make people better drivers, we need to present them with extra "near miss" accidents to keep them engaged!
But my point is we are not great at driving and I know of no one who has not at some point in their life taken chances in the form of driving when they know they are very tired and a lot of prescription drugs are clearly labeled "do not drive or operate heavy machinery" and most people I know (myself included) don't give that a second thought!
If automated systems can perform more reliably that we can, that is the measure by which these systems will be gauged. It seems they are not far from this target right now!
As far as having a person as a back up, I don't see that working out very well! Just like the "positive" test tube about every two hours, a person needs to engage their mind and if the assigned task is dreadfully boring the human mind will wander!
PS just for the record, I am 59 and have had one minor accident that was my fault; however, if I am honest, I know I have had other instances where dumb luck worked in my favor to avoid an accident!