MOUNTAIN VIEW, Calif. —
The “look Ma, no hands" moment came at about 60 miles an hour on U.S. Highway 101.
Brian Torcellini, Google's driving program manager, had driven the white Lexus RX450h out of the parking lot at one of the company's research buildings and along local streets to the freeway, a main artery through Silicon Valley. But shortly after clearing the on-ramp and accelerating to the pace of traffic, he pushed a yellow button on the modified console between the front seats. A loud electronic chime came from the car's speakers, followed by a synthesized female voice.
“Autodriving," it announced breathlessly.
Torcellini took his hands off the steering wheel, lifted his foot from the accelerator, and the Lexus hybrid drove itself, following the curves of the freeway, speeding up to get out of another car's blind spot, moving over slightly to stay well clear of a truck in the next lane, slowing when a car cut in front.
“We adjusted our speed to give him a little room," said Anthony Levandowski, one of the lead engineers for Google's self-driving-car project, who was monitoring the system on a laptop from the passenger seat. “Just like a person would."
Since the project was first widely publicized more than two years ago, Google has been seen as being at the forefront of efforts to free humans from situations when driving is drudgery. In all, the company's driverless cars — earlier-generation Toyota Priuses and the newer Lexuses, recognizable by their spinning, roof-mounted laser range finders — have logged about 300,000 miles on all kinds of roads. (Torcellini unofficially leads the pack, with roughly 30,000 miles behind the wheel — but not turning it.)
But the company is far from alone in its quest for a car that will drive just like a person would, or actually better. Most major automobile manufacturers are working on self-driving systems in one form or another.
Google says it does not want to make cars, but instead work with suppliers and automakers to bring its technology to the marketplace. The company sees the project as an outgrowth of its core work in software and data management, and talks about reimagining people's relationship with their automobiles.
Self-driving cars, Levandowski said, will give people “the ability to move through space without necessarily wasting your time."
Driving cars, he added, “is the most important thing that computers are going to do in the next 10 years."
For the automakers, on the other hand, self-driving is more about evolution than revolution — about building incrementally upon existing features like smart cruise control and parking assist to make cars that are safer and easier to drive, although the driver is still in control. Full autonomy may be the eventual goal, but the first aim is to make cars more desirable to customers.
“We have this technology," said Marcial Hernandez, principal engineer at the Volkswagen Group's Electronics Research Laboratory, up the road in Belmont, Calif. “How do we turn it into a product that can be advertised to a customer, that will have some benefit to a customer?"
With all the research efforts, there is a growing consensus among transportation experts that self-driving cars are coming, sooner rather than later, and that the potential benefits — in crashes, deaths and injuries avoided, and in roads used more efficiently, to name a few — are enormous. Already, Florida, Nevada and California have made self-driving cars legal for testing purposes, giving each car, in effect, its own driver's license.
Richard Wallace, director for transportation systems analysis at the Center for Automotive Research, a nonprofit group that recently released a report on self-driving cars with the consulting firm KPMG, said that probably by the end of the decade, “we would be able to have a safe, hands-free left-lane commute." In 15 to 20 years, he said, “literally from the driveway to destination starts to become possible."
Despite their differing goals, the approaches of Google and the car companies have much in common. They each rely on sensors to gather data about the car's environment, processors to crunch the data, algorithms to interpret the results and make driving decisions, and actuators to control the car's movements.
Most of the sensors are already in widespread use. Radar, for example, is used for features like adaptive cruise control, measuring the distance to the car ahead so that a safe interval can be maintained. Cameras are used in lane-keeping systems, recognizing lane stripes on the road so the car can be steered between them.
Digital encoders, specialized sensors that precisely measure wheel rotation, have been employed for years in antilock brakes and stability-control systems. Accelerometers have been used to measure changes in speed, particularly for air bags.
GPS devices are useful for self-driving systems, but only in giving a general sense of the car's location. More important is knowing the car's position in respect to other vehicles and objects in its immediate environment — information the other sensors provide.
“You use the sensors in the vehicle to very precisely place you locally," Hernandez said.
In the move toward more autonomous vehicles, one tendency is to integrate the data from different sensors. Camera recognition systems may be fooled by shadows, for example, thinking they are objects, but radar is not readily tricked.
Some automakers are developing a feature known as traffic jam assist, which combines the information from radar and cameras to allow hands-off driving on the highway at speeds of about 30 mph or less.
“We're taking the adaptive cruise control and the lane-keeping and bringing them together," Hernandez said.
Traffic jam assist is a step toward more autonomy, but the car is still far from self-driving; it won't change lanes, for example.
“A lot of this is getting people comfortable with the technology, showing people a benefit," Hernandez said. “The idea is the driver is always in control — the vehicle is there to help you."
Google's fleet of about a dozen vehicles adds the rooftop laser units to gather a more useful data stream than the cameras and radar systems alone can do. Laser range finders, known as lidar units, have been used by some automakers to provide distance measurements for their adaptive cruise control systems.
But Google's lidar is far more complex, consisting of 64 infrared lasers that spin inside a housing atop the car to take measurements in all horizontal directions. (Lidar systems like this are also very expensive — about $70,000 a unit — so cost and complexity will have to come down before they can be widely used.)
The units take so many measurements that, when combined with information from the radar and cameras, a moving map of the car's surroundings can be created in the onboard computer, a fairly run-of-the-mill desktop. It's a highly detailed map — the lidar can distinguish, for example, a pickup truck carrying something on a rack from a similarly sized, but boxier, delivery van.
“We like lidar because it is actually the most rich sensor you can put on a car," Levandowski of Google said. “It helps you separate out people from bushes behind them, people from each other, people from crosswalks, and it helps you make a 3-D model of the world."
Still, the key to a car being able to truly drive itself lies in the software.
“The piece that's missing is not better radars or cameras or lasers or whatever we're using," he said. “It's really the intelligence behind them."
Google's engineers tweak that intelligence based on the driving experience of the test cars. Safely coping with four-way-stop intersections was really difficult, Levandowski said, because a certain amount of assertiveness — moving into the intersection slightly to see how other cars react — is required.
“We realized there's subtle communication that goes on," he said. “Once we've come to a stop, we inch forward a bit to signal, hey, we're ready to go."
A self-driving car that did not assert itself might wind up sitting at the intersection for a long time as other cars passed on through.
Fundamentally, though, the car has to operate safely, Levandowski said, so if another car tries to enter the intersection out of turn, the self-driving car will yield.
The learning is constant. On the way back from the Highway 101 drive, for instance, an extra-long articulated bus turned in front of the Lexus, which was now back in human-driving mode because the software had been optimized for only highway driving that day. But all the sensors were still doing their jobs, so the bus showed up on Levandowski's laptop screen as a string of red dots that stretched out as the bus rounded the corner.
“Awesome bus," Levandowski said as he typed a note for other engineers to take a look.
The system constantly compares the car's map to detailed maps created by Google and downloaded to the car. Those maps provide a lot of additional information that helps with navigation, but they also help the car know when conditions have changed.
Perhaps construction barrels have just been set up, closing a lane, or a mattress or other object has fallen onto the road from a car. By comparing maps, the car knows its surroundings have changed, and it has to take some action: continue driving, alert the driver that it's time to take back control or, if all else fails, pull over to the side of the road.
The communication is two-way, so in addition to downloading Google's maps, the car can upload its map to Google. If several self-driving cars upload maps showing the new construction barrels, for example, Google can update the map it sends to other cars, letting those cars anticipate the hazard.
This connectivity is critical to Google's approach, and is one reason its system is more advanced than other efforts. (For current and planned features like adaptive cruise control, car companies have not needed to consider communication, but as they move toward more fully autonomous vehicles they will have to, experts say.)
But even Google acknowledges that its system is not there yet.
“We think it is going to be feasible for a computer to drive a car safer than a person can in the not-too-distant future," Levandowski said. “By no means are we there today. We are in the process of learning."
If and when it is introduced, there will no doubt be limits.
“What's nice about these cars is you can actually confine where they operate and how they work because they know where they are," Levandowski said.
So the system may work at first only on some highways, or in other specific situations.
“It's not going to be George Jetson from day one," he said.