This strikes me as exactly the problem of these ML/AI based vision approaches. There doesn't seem to be any higher level self protection layer on top of it to

This strikes me as exactly the problem of these ML/AI based vision approaches. There doesn't seem to be any higher level self protection layer on top of it to handle the general cases of the 1000s of things that have not been specifically trained yet/ever. The current state of Tesla AP/FSD willfully ignores the saying "Things that never happened before happen all the time". If you have ever lived somewhere like NYC you will have seen things happen on the road that can never be trained. I remember the example of FSD trying to turn left across the path of an oncoming trolley car. What is incredible was that the UI had rendered the trolley, so it knew a large object was moving towards it and yet decided to turn. Similar responses from Tesla booster of "well it was never trained on trolleys before, and now it is". What is it going to do when a car goes the wrong way down the road? What is it going to do when someone is cruising in reverse half a block to get a parking spot? We can think of untrained scenarios all day. "Big moving object closing on my trajectory" seems like a general case being missed.