The Google self-driving car has come a long way. On a demo excursion through Google’s Mountain View campus and surrounding neighborhoods today, the white Lexus self-driving test vehicle I rode in was much less of a conservative driver than I anticipated.
Sure, it followed the rules of the road, but it also accelerated into the open lane in front of us and then nudged itself around a truck that was edging into our lane so we could drive ahead without pausing.
Maybe I was kidding myself, but from my vantage point in the back seat, I didn’t feel unsafe in the least. The car braked for jaywalkers, paused when it was coming around a curve and couldn’t see whether the light in front of us was green or red, and skittered when it worried that a bus might be turning into our lane.
Some 700,000 total autonomous miles into the project, Google’s cars have never been at fault in an accident, though they’ve been rear-ended a few times. They never go above the speed limit, unless the person sitting in the driver’s seat explicitly gives permission to go up to 10 miles faster than the limit on the highway, in order to keep up with the flow of traffic.
Based on a combination of camera, laser and sensor data — GPS would be way too inaccurate — the car knows exactly where it is within a few centimeters. And yes, there’s a big red button on the center console to turn the whole system off.
But while my confidence in the system may have been boosted, I didn’t come away optimistic that self-driving cars are the way of the near future. My distinct impression was that the Google self-driving car is nowhere near ready for commercialization. There are only two dozen Lexus test models in the entire fleet. There are only 2,000 total miles of roads that have been mapped with remote-sensing Lidar so they can be driven by Google’s cars.
Indeed, five years in, Google’s self-driving car team is currently in a stage of careful tweaking. All two dozen cars are out on the street nearly every day, with two “test drivers” on board to make note of things like where the car could have rounded a curve better or, as on my drive, whether the car correctly evaluated the threat of that bus coming into our lane.
For some time, about 100 Google employees participated in a “dogfooding” test in which the cars helped shuttle them to the office. Now, the fleet is dialing in on surface-street problems like complicated intersections and construction zones.
Speaking at a Google X press event marking the opening of an exhibit about self-driving cars at the Computer History Museum, not far from Google’s headquarters, the team’s software lead Dmitri Dolgov talked about how much things have changed.
“Back in the days of 2009 when we started, everything sucked,” he said bluntly. “So you could spend half an hour hacking in the parking lot and make it 300 percent better. That is no longer the case. So we really need to drive a lot and discover challenging situations.”
But there are still some problems that Google is many miles away from solving — like driving in variable weather that could impede its sensors. Google’s self-driving cars still have trouble in rain, especially on the highway when other cars kick up water, Dolgov said. They are currently about as good as humans at driving in heavy fog. And the team hasn’t even tackled driving in snow yet, he said.
While many people who don’t trust Google’s all-subsuming ways may not feel particularly comfortable that the company is leading the way on self-driving cars, it’s not hard to come up with ways in which reliable and capable autonomous vehicles could make the world a better and safer place.
They could help reduce automobile accidents, the leading cause of death for 4-to-34-year-olds, 90 percent of which are caused by human error. They could help give elderly and disabled people more mobility, and help keep beginner drivers safe. They could make roads more efficient — apparently, for a freeway to run smoothly at peak capacity, it can currently only handle eight percent of its area occupied by cars. They could help encourage a world where cars are a shared public benefit, which could free up all sorts of resources, including parking spaces.
Okay, great. When will the mass-market version of Google’s self-driving car arrive?
Self-driving car project director Chris Urmson puts the date out at three and a half to six years — so, 2017 to 2020. Even after the convincing test drive, that seems pretty optimistic given all that’s left to do — get massively more roads mapped, cut deals with manufacturing and distribution partners, bring costs and designs down to size, get regulatory approval, address privacy and security questions and figure out the weather problems.
“If you listen to Sergey Brin, who’s my boss, back when the governor came to Google to sign self-driving legislation, he said within five years, and that was a year and a half ago, so that’s one timeline that we’re working against — but that could be trumped by safety,” Urmson said. “The other timeline is that I have a 10-year-old son, and if you look at the statistics for teenage drivers, they’re terrible. So I’ve got six years.”
As Dolgov described it, the crazy thing is that all of the millions of measurements and computations per second come down to exactly two things. Two numbers, really: How much to adjust the throttle and steering.
“The challenge, of course, is there are many combinations of those two numbers that you don’t want the car to do. We need to find the best answers, and we need to do that many, many times per second if we are ever going to get to that point where you are playing dominoes in the car.”
This article originally appeared on Recode.net.