That last bit is really a great example, though- a computer is not going to be able to say 'well, it looks like I should go off onto the shoulder and risk scraping a guardrail/curb (lesser evil in general) rather than endure this massive collision that awaits me on the road.'
It also wouldn't be able to make a decision regarding what would be the softer, more forgiving thing to hit, or what action would endanger nearby pedestrians the least.
I know extremely well how bad your typical driver is. But my point was kind of that the driverless car would not have any leeway for mistakes. As soon as a 'robot' murders a child and leaves no one to blame other than the carmaker, people aren't going to want them anymore. And the manufacturers won't want to build them.
A computer could absolutely make those decisions. I'd wager it would, overall, do a far better job of it than humans due to faster reaction times and better spatial awareness. The computer decision tree would look like this:
Stopped car ahead -> Do I have time to stop? -> If yes, stop and notify cars behind so that they can start braking as well -> If no -> Is there an open lane or shoulder in this traffic direction? -> If yes, go there -> If no, apply full brakes to minimize impact
Cars can already detect pedestrians, so deciding to not ramp up onto a busy sidewalk to avoid an accident wouldn't be a terribly difficult task. It would be able to make these decisions in less time than it would take a human to check their mirror.
I also don't trust whoever would have control of the central "brain" that monitors/controls the cars. Yes, there would have to be a central brain to make everything work so perfectly as some of you claim it would operate, in order to synchronize everything.
You could probably get better results with centralized traffic planning, but you could get 90% of the way there by giving each car current, high-quality traffic data plus a database of historical data for heuristic route planning.
