I also see questionable engineering decisions. For example, some reverse engineering showed that there was an overlay map of problematic stop signs and road signals that was downloaded per-road on all navigation routes. Surprise surprise, as that data got out of date, the car would randomly slam the brakes on as the overlay data told it there must be a stop sign, when in fact it was a temporary sign from construction months ago.
If you're aiming for end-to-end neural networks, you shouldn't have hard overlay facts like this. Ideally you get all the behaviour you need from curating training data, but if you can't do that the most you should have location specific info-vectors that nudge the network into making the right decision if needed. Such info-vectors can be used for all kinds of things, like "this city has aggressive drivers", "this street has a blind corner", "there's a dog at this house who loves to run under wheels", "this country drives on the other side of the road", etc.
If you're aiming for end-to-end neural networks, you shouldn't have hard overlay facts like this. Ideally you get all the behaviour you need from curating training data, but if you can't do that the most you should have location specific info-vectors that nudge the network into making the right decision if needed. Such info-vectors can be used for all kinds of things, like "this city has aggressive drivers", "this street has a blind corner", "there's a dog at this house who loves to run under wheels", "this country drives on the other side of the road", etc.