[jcl]

On the other hand, humans also have the common sense not to (e.g.) follow the tracks of a car that has spun off the road.

I'd expect logic at that level might be more difficult to train for than the relatively low-level visual task of identifying and following tracks. But I guess existing self-driving systems already need to work at that level to predict the actions of other cars and pedestrians, so presumably the current data-driven training techniques would also handle it OK? (Assuming a lack of pranksters with shovels...)