I've been following the development of Tesla and Waymo's 'self driving' systems.
Tesla seems to make 'two steps forward, one step back'. The seem to have definitely 'overfit' to parts of the California area - the 10.3 beta version seems to have partially addressed that - the 10.2 beta had a distribution to a much wider geographic distribution than prior betas (which limited to access to less than 100 people outside of Tesla). So road layouts different from the cities it drives in most can confuse it. Also 10.2 had very aggressive acceleration when pulling out from lights, and did California rolls - 10.3 much less aggressive (more natural speed), and whether to complete stop or not is configurable.
Some night driving and driving directly into a sunrise/sunset seem to give the system difficulties, also areas where they don't have good 'HD maps' (They do 'HD' mapping via collecting drive data, then having the company TomTom process the data).
It definitely does unpredictable behavior occasionally and is not ready for distribution from people who aren't vigilant, but on areas that it has 'experience' it does pretty amazing.
For instance here is a frequently tested route of a curvy drive
https://www.youtube.com/watch?v=CjYNsGcQD3AThis is a frequent route - it does reasonably well but a bit too aggressive
https://www.youtube.com/watch?v=Ui9qjDgVXcoThis is a rather scary fail, when it looks like it thinks the railroad tracks road is the left hand lane, and doesn't seem to 'see' the light pole and an inattentive driver might have ended up in an accident.
https://youtu.be/bbyNg9kYEq4?t=344
