Localization is primarily based on visual registration, i.e. matching the current surroundings to the closest data in its map. Lane markings are based on map data and what it's able to see in real-time.
Maybe similarly how humans do it.
Humans do it super badly in a heavy rain/snow though. We're basically blind and just toodle along following the guy in front and hoping for the best.
A machine driver should not accept these conditions.
On the contrary, sticking strictly to the lane markers when everyone else is blindly straddling lanes seems the worst of all worlds.
Right, but I’m arguing that the correct behaviour for everyone is to pull over and wait for conditions to improve — robots and humans alike.