Re: Zero accidents?
It's one thing promising perfect performance on "business as usual" activities. The problem with that is, accidents aren't business as usual. They're edge cases and that's just what software has always had trouble in handling. And these are not just simple edge cases such as off-by-one that can be tested for. They're going to be "we never saw that coming" events. They're going to be events that require much more processing than normal to deal with an unexpected set of circumstances.
If a designer reckons there might be enough processing power to cope then maybe the "run down the lone pedestrian" option gets hard-coded as a would-be damage limitation short cut. And then that gets triggered by some freak set of circumstances when an accident wasn't threatened and the car goes out of its way to run one down.
Another issue is certification. That's going to be a difficult one to test. Will there be a temptation to code to the test? Remind me, who was it who wrote the code for the VW emission control?