Google's self-driving experiment heads towards next phase

MOUNTAIN VIEW, Calif. -- Google's self-driving car has never driven in the snow, gets puzzled by parking lots and cannot comprehend the hand signals from a traffic cop in an intersection.

For all its cameras, lasers and sensors, the car of the future still can't do all the things that human drivers can.

And yet, Google has come so far so fast with its 5-year-old experiment in autonomous driving that members of the self-driving team are now speaking more confidently than ever before about the next phase of its ambitious research project: moving the technology out of tricked-out test vehicles and into real-world cars and trucks.

Google executives talk regularly with multiple automakers about the technology, and they're thinking hard about the best way to deliver a product that makes driving safer and traffic jams less tedious, Christopher Urmson, the director of the self-driving cars project, told reporters during a rare media briefing last week near Google's corporate headquarters here.

"It has to be at a price point where the value to the customer exceeds the cost to the customer," Urmson said of the eventual product. "We're working on that."

In Google's eyes, the imperative is strong. Each year, more than 30,000 people die on U.S. roadways from accidents mostly caused by human error. Ron Medford, who was second in command at the National Highway Traffic Safety Administration before becoming director of safety for the self-driving cars project in 2012, said he expects autonomous cars to be one of the three biggest lifesavers in the history of auto safety, along with seat belts and electronic stability control.

After last week's media briefing, Google drove dozens of journalists on 25-minute loops of Mountain View, during which the car confidently followed a preset route with virtually no need for human intervention. (See story, Page 30.)

The company made clear that it still has big technological challenges ahead, including refining the maps and software that have allowed its modified vehicles to rack up more than 700,000 miles in autonomous mode since 2009. Google also must win the blessing of safety regulators and the trust of a skeptical public.

A fully autonomous vehicle will not be offered "until it's safe and ready to be on the road," Urmson said.

"We think about that constantly. We want to make sure that when the technology comes to market, that you trust it enough to have it out there."

But each step forward in Google's project widens the path for automakers to shift more control from driver to car, a trend that began decades ago with automatic transmissions, power steering and cruise control and now encompasses features that can intervene to prevent a crash, such as electronic stability control and automatic braking.

More automakers are now looking at making the jump to a fully autonomous car. Volvo recently announced plans to put 100 self-driving cars on public roads in its hometown of Gothenburg, Sweden, by 2017. Mercedes-Benz said last year that a modified S-class sedan drove 62 miles from Mannheim to Pforzheim, Germany, on city and highway roads. Nissan, for its part, pledged to have a fully autonomous car ready for deployment by 2020. General Motors, moving step by step, says by 2020 it will launch a feature called Super Cruise that can take control of the steering and pedals once a driver is locked into a highway lane.

Google's prototype self-driving vehicles have become a familiar sight here. Originally, the fleet consisted of modified Toyota Priuses, and, since 2012, modified versions of the Lexus RX 450h.

The heavy testing has helped Google work out kinks in everything from mapping to software to the costly sensors that provide a 360-degree view of a car's surroundings.

The vehicles start off with a combination of off-the-shelf cameras and radar sensors, like the ones automakers use for features such as adaptive cruise control, lane departure warning and automatic emergency braking.

The one big difference is the spinning, roof-mounted laser turret, which Google gets from Bay Area technology firm Velodyne Inc.

Underlying all the systems is a software combination of GPS and 3D maps that depict the world in detail that goes well beyond Google Earth or Google Street View. Google says it has mapped out about 2,000 miles of roads at that level of detail, which includes everything from curb heights to lane markings to traffic lights.

Andrew Chatham, mapping lead for Google's self-driving cars project, said the systems work together to answer four questions: Where am I? What's around me? What will they do? What will I do?

The answer to the first question comes from looking at GPS and the 3D maps. "We tell [the car] what the world is expected to look like when it's empty," Chatham said. "It's the job of the software to figure out how the world is different from that expectation."

Ultimately, the task of controlling the car is fairly simple, says Dmitri Dolgov, the software lead at Google. It basically requires two inputs, one for acceleration and braking and one for steering. But to get to that point, Google has had to write an enormously complex program to do what human drivers do in their heads: predict how cars, cyclists and pedestrians will behave and be ready to react to every situation imaginable. Google's cars are programmed to handle thousands of driving situations, including crossing railroad tracks and creeping into a crowded intersection to make a left turn.

Google's early tests in 2011 and 2012 focused mostly on freeways, which are normally free of cross-traffic and pedestrians and far more predictable than surface roads. In early 2013, tests transitioned to urban driving, as Google mapped out hundreds of secondary and tertiary streets across Mountain View, Chatham said.

Still, Google has mapped only about 2,000 miles of roads in total -- merely as much as it needs to refine the technology, Chatham says. To commercialize it, Google would need far greater coverage. The United States has 4 million miles of public roads. "We really have our work cut out for us," Chatham said.

Google engineers haven't said much yet about the business model for the company's autonomous car technology. It could try to sell or design its own car but is seen as more likely to partner with manufacturers to have its equipment installed.

Google co-founder Sergey Brin set a goal in late 2012 of having the technology commercialized within five years.

Urmson said last week he has his own timeline in mind: going to market by the time his 10-year-old son starts to drive, six years from now.