The book was better
Google Car’s first at-fault accident.
Honestly, I expect things like tumbleweed and dust storms will be a big problem for the early generations. But once some event happens a few times, if it proves to be a real nuisance, the engineers can start training the networks for it.
An anecdote, one of the early test models had a problem with riders of fixed gear bicycles. See, the system had been trained to recognize cyclists, but only on "freewheel" setups. A typical freewheeler will stop at an intersection and put their legs down. "Fixed gear" folks, however, I guess often keep their feet on the pedals and kinda do a short forward-back cycles to keep balance while waiting for the intersection to clear.
The cars, of course, had no idea what was happening, because the bike kept "lurching" forward as if it was going to go into the intersection. So they waited. Forever.
Until they trained the network on that. Now it can sense that behavior and make a correct decision.
Likewise for accident avoidance. This becomes a "global" (more "within local area," if taken literally) control/optimization problem. But just as a person can direct traffic, a computer can. Of course, this would require coordination, communication, and a number of other solvable problems.
My point, the problems are real. They are also solvable.