Bob mentioned the robo-taxi in a post below so I thought I'd start a new thread with my  thoughts.
It seems to be much, much, much harder than anticipated. Go back in time a few years. It was publicized that Teslas with HW2 were hardware capable four years ago and it would be working by the end of the year. Since then, the hard coded C++ algorithms have been scrapped and a new approach involving end to end machine learning has been adopted. HW2 has been replaced with HW3, that's been co-opted by HW4 and even that isn't up to the task, so HW5 with more compute and better sensors is on deck.
I remain skeptical for a few reasons: None of the solutions, so far have made use of short term memory. The system may see a raster representation of objects, vectorize it, and recognize it by painting a block diagram on screen. If it disappears from view for only a second, the system forgets it was ever there and has to go through the whole process of raster to vector conversion and recognition. This becomes important when driving on hilly 2-lane roads. If I'm driving, I can see a mile ahead so if I drop into a dip that obstructs my vision, I still know there is an oncoming car or maybe not so I can start passing as soon as I top the hill. FSD has no such opportunity because it has no memory of what it has seen just seconds before. It is always dumb as a rock when history is concerned. It has no ability to do anything other than react and until it is able to remember and be proactive it will struggle in some environments where even an old man can out-think it. 
And then there are instances where even great cameras will not be able to overcome some of its flaws. Montana speed limit signs continue to confuse the system. I frequently drive stretches of road between Wolf Point and Glasgow and between Wolf Point and Circle. Without fail, the system misses several of the 70 MPH signs along the way. I'm right next to them and still they don't get interpreted correctly and It's not because of poor cameras because it seldom misses the slow speed limit signs. 
There is another stretch of I-94 where crack sealing has been done right along the centerline. If there is just a small patch of white paint that hasn't been obscured by the bright black tar, the system interprets a center line. But there is a stretch of half mile or so where the white paint has been completely covered by black tar. An average driver will have no trouble figuring out that the black tar is the center line. The auto-pilot is completely confused there and the car will proceed to occupy both lanes. And to top things off, it does this every time I drive that stretch! It has no memory of prior screw ups.