The car was unsure of itself quite a few times, stopping in illegal and extremely dangerous ways.
I'm not seeing facial expression analysis to increase probability accuracies for questions like do they see you and do they see other vehicles to which they may or may not appropriately react, and how much attention is on new ideas like turning right or left.
It was sped up so you can't see blinkers or other key signals.
I can't tell from this display if they calculate speed, attitude, driver function of each car, to predict their possible motivations.
I can't tell if they are willing to cross a double yellow line to avoid running over people. Do they cause stopping accidents maybe killing pedestrians in order to not cross double yellows?
I can't tell if it honks at dangerous situations.
I can't tell if it quickly starts to slow down slowly, watching rear cars to see if they react, and continue to monitor the rear cars progress at stopping while increasing stopping speed, until safely and comfortably stopped.
I don't see multilane analysis, lane path probabilities, and other things about driving.
But, it's looking better than the original video.
The less they hand code and the more they generalize rules, the better: never sudden changes unless emergency; try to figure out what will happen, and plot course to probable destinations (skip stop at nearby store if rush hour traffic increasing; etc.) through this without hitting anything bad; attempt to follow laws. General rules like that would be better than hand coding each little thing. Pro race drivers and truck drivers could test its knowledge to make certain it is thinking correctly. Throw in there also chofeurs(sp) and extreme driving teachers to round it out.