Humans do an unbelievably bad job at this, too, like all other tasks behind the wheel. Humans essentially just follow the tracks of the last car, creating ad hoc lanes. Computers can do this just as well.
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...)
The smart ones pull over and turn on the emergency flasher. Then they check to make sure they have lots of gas, blankets and water, just as they prepared.
The dumb ones slow down to 45 mph and stare into the white blankness looking for clues about curves.