I CAN’T TELL from the back seat who’s in control of Google’s autonomous Lexus. The SUV is prowling the streets of Mountain View, California, but it’s only after lead engineer Dmitri Dolgov flings his hands into the air that I realize he may be in the driver’s seat, but the computer is in control.
After a flurry of pronouncements from companies racing headlong into autonomous driving, Google’s venerable project feels like an afterthought. After all, Ford CEO Mark Fields promises hundreds or thousands of autonomous taxis by 2021. Lyft co-founder John Zimmer says most customers will tool around in robo-cars within five years. Uber has a small fleet of self-driving cars ferrying people around downtown Pittsburgh, Pennsylvania, right now.
All of this makes it easy to forget Google is the old hand here. It started developing the technology in 2009. Granted, the company has never said anything remotely useful about when it might be ready, where it might be used, or what people will do with it. Instead, it keeps testing and refining and testing some more, using a fleet of nearly five dozen cars in Mountain View, Phoenix, Austin, and Kirkland, Washington.
That fleet just hit a remarkable milestone: Those Lexus SUVs and dorky prototypes rolled the collective odometer over two million miles of autonomous driving on public roads, 90 percent of them on city streets.
There’s nothing inherently significant about the number, but it’s a reminder Google remains the veteran player in a field full of rookies. In those two million miles, a robo-car caused just one minor crash. The technology experienced a glitch that required human intervention just 341 times during one 423,000-mile stretch. That’s not quite as foolproof as carbon-based lifeforms, who tend to crash once every 238,000 miles, but it’s not bad.
That experience matters, because Google is working on the hard stuff now. “It’s fairly easy nowadays to get the first 90 percent of the problem,” says Dolgov, the program’s head of engineering. That last 10 percent requires teaching the car to handle edge cases, from that couple riding unicycles to the woman in an electric wheelchair chasing a duck with a broom. (No, really. Those two things actually happened.)
When the car encounters something new or falters by, say, taking a curve too fast or driving too cautiously, the human in the passenger’s seat makes a note in her laptop. That info goes back to the lab for “fuzzing”—that’s when Google’s driving simulators take a moment, tweak its elements, and teach the software to handle nearly endless variations. You think two million miles in eight years is a lot? Google simulates three million miles every day.
That experience shows. When I rode around in Uber’s self-driving car last month, the human engineer took control every few minutes for stuff like avoid an endless wait behind a double-parked truck. I never felt like my safety was at risk, but it was clear the car wasn’t ready to go it alone.
During my 30 minute tour of Mountain View, Google’s car never needed a hand. Granted, the Silicon Valley suburb is less complex than the City of Bridges, but the autonomous Lexus drives like an experienced human. At a four-way stop, it waited to see if an approaching cyclist would roll through the intersection (he did). When it spotted a pedestrian crossing some 30 yards ahead, it braked. When a red light turned green, it waited a full second before hitting the gas to ensure sure the way was clear.
Perfection eludes the machine. As it entered a bottlenecked street, it jerked to a stop. Twice, it tapped the brakes for no discernible reason. Passing a parked truck on the right, it swerved to the left.
Any driver’s ed teacher would be proud.
Click Here To Watch how Google has gone this far.