Moveable

California Is a Terrible Place to Test Autonomous Cars

Self-driving cars need testing in the rain and snow, so why is everyone in the Bay Area?

Tracey Lindeman

Tracey Lindeman

A self-driving car in Mountain View, California. Image: By Runner1928/Wikimedia Commons

There are now 45 companies testing self-driving vehicles on Californian soil, mostly in the Bay Area—a place known for having T-shirt weather nine months out of the year.

From Volkswagen to Samsung to Lyft, transportation and technology companies are trying to figure out how to optimize self-driving cars. Loaded up with sensors, cameras, and other tech, the cars shuttle around San Francisco, Mountain View, and Palo Alto trying not to hit anyone or anything (with varying degrees of success).

According to these companies’ disengagement reports from 2016, suboptimal weather conditions are one of the recurring reasons for test drivers having to switch off the autonomous drive.

Alphabet Inc.’s autonomous-car subsidiary Waymo recently published a Medium post where it flaunted its four million hours of autonomous-car testing, done in 23 US cities—all of which are decidedly winterless, save for a test done at Lake Tahoe in late March.

But how much can we ever expect the general public to buy into self-driving cars if they haven’t trudged through rain, slush and snow?

Fred Tung, a computer-vision researcher in British Columbia, drove around Vancouver on a rainy night to try and find out.

“The snow presents challenges you will never experience in sunny California”

Tung didn’t have an autonomous car at his disposal; rather, he used a dashcam to record his drive. Then he returned to the lab to figure out how well the 720p camera was able to detect objects in low-visibility settings.

The answer: Not well.

“In general it’s a very difficult problem. Up to now in this field we’ve been just trying to get [self-driving cars] to work in good weather and good lighting,” Tung told me in a phone interview.

As he expected, factors like rain, glare, and pedestrians dressed in dark colors made it difficult for the camera to understand what it was seeing and parse the scenes accordingly. He detailed the results in his study, “The Raincouver Scene Parsing Benchmark for Self-Driving in Adverse Weather and at Night.”

Fred Tung collected this image while driving around rainy Vancouver with a high-definition video camera strapped to his dash. Image: Fred Tung

“Safe and reliable operation in adverse conditions is necessary for the mass market adoption of self-driving vehicles,” his study concluded, calling for cameras to be paired with different sensors to better interpret and adapt to real-time road conditions.

The most common sensor, LIDAR—which uses pulsed laser light to determine proximity and depth of objects near a self-driving vehicle—is still being tested for its usefulness in inclement weather.

One of the problems with LIDAR and other mounted sensors is that they can easily get covered in snow, affecting their ability to “see.” LIDAR also gets confused when its lasers get absorbed by snowflakes and water droplets, said Huei Peng, the director of Mcity in Ann Arbor, Michigan, a simulated urban environment for testing autonomous vehicles. “The snow presents challenges you will never experience in sunny California,” he said.

Read More : We Talked to a Driver Who Disguised Himself as a Car Seat

Another issue is that high-performance LIDAR is super heavy and expensive, costing around $70,000 per unit, one of the reasons why Tesla has come out against using it in its cars. (Increased demand has helped lower LIDAR prices recently, but they still cost $8,000–$30,000 each depending on the model.) “Until LIDAR becomes cheap we’re going to have to think about using cameras,” said Jim Little, Tung’s PhD supervisor at the University of British Columbia.

Dedicated short-range communications, too, could pick up some of the slack by letting cars speak to each other and everything around them (known as V2V and V2X, respectively), Peng pointed out.

Equipped with all of these tools, some makers of self-driving technology have heeded the call of the north to conduct testing in colder climes.

At the end of November, Russian software company Yandex published a video showing off how its self-driving car masterfully handled snow-covered roads.

Ford Industry's first autonomous vehicle tests in snow. Video: Ford Media/YouTube

Waymo recently announced a partnership with Chrysler that would bring it to Michigan for some wintry testing. Also in Michigan, Ford Motor Company—which is testing self-driving cars at Mcity—said last year that it’s using ultra-detailed road mapping to overcome some of the weather-related limitations of self-driving cars.

However, one of the issues with mapping is that if that if an actual road changes in any way—newly painted lines, road work, potholes, detours—a new map will likely need to be created. “Maps are only as good as how often they’re updated,” said Peng of Mcity.

In Pittsburgh, Uber lapped up ample media attention when it launched a fleet of self-driving cars in September of 2016 with the intention of testing them throughout the winter, though it remained customarily tight-lipped on how successful those tests actually were. The company had previously demanded that Pittsburgh prioritize its test routes for snow removal. It also benefitted from a mild winter with lower-than-average snowfall in 2016–2017.

Meanwhile, Canada’s Automotive Parts Manufacturers’ Association (APMA) manages a new demonstration zone for connected and autonomous vehicles in snowy southern Ontario. APMA president Flavio Volpe said automakers and other companies have “expressed great interest” in testing out their technologies there: “You get to see the supplier’s product in action, in real life, in 12-month weather.”

Get six of our favorite Motherboard stories every day by signing up for our newsletter.