Then Google can teach the cars to do just that. You seem to think there are things these cars can't even conceive of, which is patently nonsense as they are trained by humans. If there is a great way to behave in a certain situation, the car can be trained to do just that, but with its far superior reaction times.
Try your test with human drivers and see how many toddler dummies you get through.
And then the human driver gets lost in thought (or tired, or angry, or distracted, etc. etc.) and crashes anyway, as they are intrinsically terrible at driving.
You should heed your own advice - if anecdotal stories are not evidence, what does that make your ass-delving hypothetical situations and guesses?