That's actually what I figured this article was about, based on the title.
Also, the distribution of the problems themselves are going to be very time sensitive. When the car yells "Oh crap, I need a human!", you're looking at at least 5 seconds for a human to respond. Even ignoring the "take a number" problem, which pushes that number arbitrarily high during times of high load. Whatever problem resulted in the car yelling "I need a human!" is probably resolved one way or another before the human gets there, because physics isn't waiting. If the car persistently needs a human, the answer is not to remote-drive the car but to stop driving entirely, so it's only the transient issues that matter.
Basically, I'm not clear on what problem this solves.
Lately they've been turning their attention to more pressing needs, like learning how to help the handicapped drive.
If that sounds absurd, it's not. If you're trying to take advantage of the computer driving to run up against your safety margin, that has further effects. Sure, most of the time the computerized cars could in fact drive mere inches from each other's bumpers, but in an emergency situation this means you have no margin. You can't both spend your margin and have it when you need it.
If anything I expect robotic cars to end up being more cautious than humans in the end. Automated cars will be able to precisely calculate the risks of given driving behaviors just by their nature, and I seriously doubt that car companies are going to have any stomach for permitting their cars to engage in complicated gas-saving car behaviors that will provably be more dangerous than driving safely, which I suspect will look rather a lot like what highways do now, except with less variance in the speeds.
One of the things I'm really curious about is whether during a winter storm the cars will simply refuse to drive. That will be an interesting cultural change up here where snow is actually a thing.