Autonomous Cars Can Now Drive on Snow-Covered Streets

If you dread having to drive down a snow-covered Canadian road, just be thankful you're not a self-driving car.

Compared to computers, human drivers have it easy in the winter. Even if the snow is covering up important markers such as lane lines, humans are pretty good at synthesizing what visual information remains and making educated choices about where to go or how to drive.

But under similar conditions, self-driving cars run into trouble because they can't "see" these markers, which they depend on under normal circumstances. Google's self-driving car project director Chris Urmson admitted as much himself last year when he said that snowy conditions are a problem for Google's fleet of experimental self-driving cars.

Researchers at the University of Michigan's Mcity—a fake town built just for self-driving car testing—announced on Monday that they've come up with a way to get around this limitation, however, and that the autonomous cars they've tested can now navigate on a snow-covered road with ease.

"Autonomous cars rely on the appearance of the roadway surface itself—road paint, and even imperfections in the roadway itself—but in the snow, all of those disappear," said Edwin Olson, a researcher at Mcity. "We decided to use the 3D structure around us, things like road signs and buildings, as navigational landmarks. And the nice thing about those is that they don't get covered up by snow, they don't get repainted."

Watch more from Motherboard: The US Military's Robot World Championship

One important thing to note is that this only solves one small part of the problem with driving on a snow-covered road. Other issues, like avoiding black ice or pedestrians when sensors are obstructed by flurries or fog, still need to be figured out. But it's a step forward.

When driving down a snow-covered street, for example, a self-driving car may not be able to see the lanes, but it can use its sensors to detect landmarks such as buildings and other objects nearby. The car compares these landmarks against a pre-made map of the area—collected by scout vehicles that have previously driven the route—and uses the results to orient itself on the road.

"What we're looking at is a fleet of survey vehicles going in and doing the initial mapping," Olson said. "This mapping data gets vetted and filtered, even by humans if need be, and that data is what gets streamed to the car as the car is driving around."

Most autonomous vehicle approaches, including Google's, use pre-made maps, Olson told me, in lieu of advanced AI that can calculate its position without any help. Cars may one day be smart enough to safely do this, but until then, they'll need a little help to get out onto the road.

In the meantime, Olson emphasized that the work at Mcity is a good start to tackling these issues, but more testing awaits. The only things standing in their way?

"Weather and lawyers," Olson said.