If you zoom out, you'll see that there's a method to the madness but a lone driving car that's not a part of a fleet doesn't have such luxuries.
I don't think the problem is solely image processing and machine learning. I just don't get self driving and how we're supposed to wow the logic. You have to be aggressive to drive in a city or people will constantly cut you off for hours. This means you will hold up traffic making the situation worse for everyone. However, you can't be too aggressive. Imagine the damage we will cause to the "optics" of self driving if we run over (at a slow 10 mph) stubborn people who cut us off. That will just look like murder.