The First Look at How Google's Self-Driving Car Handles City Streets
This post originally appeared in the Atlantic Cities.
The first rule of riding in Google’s self-driving car, says Dmitri Dolgov, is not to compliment Google’s self-driving car. We’ve been cruising the streets of Mountain View for about ten minutes. Dolgov, the car’s software lead, is sitting shotgun. Brian Torcellini, the project’s lead test driver (read: “driver”), is sitting behind the wheel (yes, there is a wheel). He is doing no more to guide the vehicle than I’m doing from the backseat. I have just announced that so far the trip has been “amazingly smooth.”
“The car knows,” says Dolgov.
He means I have violated some code of robotic superstition, calling the contest too early. Or maybe he means my praise serves no function here. If I can tell how well the car is driving itself, so can the car.
Google’s self-driving car project began in 2009. The vehicle’s early life was confined almost entirely to California highways. Hundreds of thousands of test miles later, the car more or less has mastered the art—rather, the computer science—of staying in its lane and keeping its speed. So about a year and a half ago, Google’s team shifted focus from the predictable sweep of freeways to the unpredictable maze of city streets. I was invited along as the first journalist to witness how the car is handling its new urban lifestyle.
Over the next few minutes the autonomous vehicle makes several maneuvers that someone less privy to Dolgov’s first rule would have been tempted to compliment. We go through a yellow light, the car having calculated in a fraction of a second that stopping would have been more dangerous. We push past a nearby car waiting to merge into our lane, because our vehicle’s computer knows we have the right-of-way. We change into the right lane for a seemingly pointless reason until, a minute later, the car signals a right turn. We go the exact speed limit because maps the car consults tell it this road’s exact speed limit. The car identifies orange cones in the shoulder and we drift laterally in our lane, to give any road workers more space.
Between you and me: amazingly smooth.
Equally amazing is that people around us are going about their daily lives. I’d read that drivers tend to gawk at the Google car from their own cars, but that is not the case today. At one intersection I look at the cars flanking us. The driver to our right finds her cell phone more fascinating than we are; the driver to our left rests his head in his palm, and may or may not be falling asleep. There is a banality to vehicle autonomy in this place.
It can’t be that they’ve missed us. If the spinning bucket suspended by four metal arms on the roof doesn’t give us away, the words “Self-Driving Car” on the rear bumper should. We’re in a white Lexus RX 450H, part of a fleet of about two dozen prototypes, all of which now spend most of their time on surface streets. The bucket spins ten times a second, emitting 64 lasers that generate 3D information on objects all around us; the car also has radar that bounces 150 meters or so in every direction to perceive things a human driver never could. The Lexus’s interior is standard with the following exceptions: a camera facing out from the windshield capable of reading traffic lights, street signs, etc.; an ON and OFF button on the steering wheel to engage or disengage “auto” mode; a driver’s side display panel showing our speed and position; and a big red button on the wooden console—a kill switch they’ve never had to use.
“Every robot has a big red button,” says Dolgov.
Dolgov is holding a laptop running a map that effectively displays what the car is “seeing.” There is a comment box on the screen where he can record notes should something of interest occur during the ride. Right now he is not recording any notes. “Not much interesting stuff is happening,” he says. I had actually been promised ahead of time that “interesting things” would happen during the ride, so I could feel a bit misled at this moment. Except I’m riding in a car that’s driving itself through a city so amazingly smoothly that people around us are falling asleep.
In that sense this uninteresting ride feels profoundly, even unimaginably, interesting.
• • • • •
The head of Google’s self-driving car project is Chris Urmson, a tall man with tousled blond hair and a boyish grin to match an idealistic spirit. We met at a Google X building just before my test ride. Google X is the company’s tight-lipped (but loosening) innovation lab that both oversees and emerged out of the self-driving car project. It is known for impossibly lofty goals with a sci-fi twist; its director, Astro Teller, is officially titled Captain of Moonshots. Urmson shares a resistance to incremental advance.
“You make so much more progress when you’re thinking about changing the world rather than making this minor delta improvement on something,” he says. “You can get fired up in the morning.”
Urmson came into driverless cars like so many in the field: via three autonomous vehicle challenges held by DARPA in the mid-2000s. The first Grand Challenge, in 2004, was a legendary disaster. Urmson was part of a team from the robotics institute at Carnegie Mellon led by former Marine William “Red” Whittaker. They made the contest’s best showing despite traveling just 7 of 150 miles before getting stuck in an embankment. “Almost literally burst into flames,” says Urmson. At the next Grand Challenge, in 2005, they placed second and third, losing to a Stanford group led by Sebastian Thrun, who later started Google’s self-driving program. Urmson’s team did win the 2007 race—an “Urban Challenge,” notably, through 60 miles of a city environment.
He came to Google in 2009 to develop the self-driving car because it felt like something “that might change the world.” Urmson knows the statistics on metro area congestion. Americans spend 52 minutes a day commuting, he says, which works out to 4 percent of their lives. (“If I could give you 4 percent more life, you’d take it.”) His bigger goal is safety, and he recites these numbers, too: 33,000 people a year die on U.S. roads; car crashes are the leading cause of death for people age 4 to 34; at least 90 percent of collisions are the result of human error. “So this is kind of a big deal,” he says.
After accomplishing two baseline goals in its first 18 months—one, to drive 100,000 miles on public roads; the other, to complete ten 100-mile courses on challenging routes throughout California—the Google car spent the next couple years conquering freeways. That seemed a “simpler problem” to tackle first compared to city streets, says Urmson. Yes, higher speeds make the potential cost of any mistake that much bigger, but the fundamentals of freeway driving are pretty easy for programmers to model. Cars move in one direction, making minor adjustments to speed and position.
“To grossly simplify it,” Andrew Chatham, the project’s mapping leader, later tells me, “you follow the curve and don’t hit the guy in front of you.”
It’s not just that surface street driving is far more complex than freeway driving, it’s also unpredictably complex.
Cars move at slower speeds on city streets, but the number of variables are almost endless, and they require vigilant attention in every direction. There are tight lanes and traffic lights, pedestrians and cyclists, oncoming cars and double-parked trucks, unprotected turns and unexpected road work—the external elements are infinite, and configured differently each trip. So it’s not just that surface street driving is far more complex than freeway driving, it’s also unpredictably complex.
Take the problem of crosswalks at intersections. Sometimes pedestrians wait for the crossing signal and walk inside the lines. But sometimes they ignore the signal and cross as they please, and sometimes they’re just waiting on the curb for a friend and don’t mean to cross at all. Early on, the Google car had trouble categorizing these varying intentions and deciding how to respond. Now it’s graduated to subtler problems, like spotting a pedestrian who might be standing behind a utility pole at the corner.
“It’s the rarer and rarer situations we’re working towards,” says Urmson. “The complexity of the problem is substantially harder. But basically over the last year we’ve come to the conclusion it’s doable, and that this intuition we had about making a vehicle that was fully self-driving was correct. That it was possible. That we actually think we can make one that really is safer than human driving.”
• • • • •
An interesting thing has happened in the car. We are in the left lane on Mountain View’s West Middlefield Road when some road work appears up ahead. A dozen or so orange cones guide traffic to the right. The self-driving car slows down and announces the obstruction—”lane blocked”—but seems confused what to do next. It won’t merge right, even though no cars are coming up behind us. After a few false starts, Brian Torcellini takes the wheel and steers around the cones before reengaging auto mode.
“It detected the cones and it tried to go around them, but it wasn’t confident,” says Dmitri Dolgov, typing at the laptop. “The car is capable of a lot of things, but unless it’s absolutely sure that it can handle some situation well, it will err on the conservative side.”
Boiled down, the Google car goes through six steps to make each decision on the road. The first is to locate itself—broadly in the world via GPS, and more precisely on the street via special maps embedded with detailed data on lane width, traffic light formation, crosswalks, lane curvature, and so on. Urmson says the value of maps is one of the key insights that emerged from the DARPA challenges. They give the car a baseline expectation of its environment; they’re the difference between the car opening its eyes in a completely new place and having some prior idea what’s going on around it.
Next the car collects sensor data from its radar, lasers, and cameras. That helps track all the moving parts of a city no map can know about ahead of time. The third step is to classify this information as actual objects that might have an impact on the car’s route—other cars, pedestrians, cyclists, etc.—and to estimate their size, speed, and trajectory. That information then enters a probabilistic prediction model that considers what these objects have been doing and estimates what they will do next. For step five, the car weighs those predictions against its own speed and trajectory and plans its next move.
That leads to the sixth and final step: turning the wheel this much (if at all), and braking or accelerating this much (if at all). It’s the entirety of human progress distilled to two actions.
The map on Dolgov’s laptop screen offers the best visual window into the car’s mind’s eye. Take the screenshot from one of our right turns (below). The baseline image is the detailed area map in grayscale. Layered atop that are objects identified by the car’s sensors, depicted in colorful geometric boxes: purple for vehicles, red for cyclists, yellow for pedestrians. The red and green ladders are objects that have an immediate impact on the car’s speed; in this case, though the traffic light is green, pedestrians prevent a turn, as does a cyclist coming up on the right—in a spot a human driver might easily miss. The flat green line shows the car’s planned route.
Dolgov logs the road work incident in the computer. He explains that feedback from the driving teams is critical to the car’s development. “Every disengage has a severity associated with it,” he says. “That was not the end of the world. We would have gotten through the cones. But it was a problem. Once we go back we’ll pull the disk out of the car. We’ll import the log from this run. This will get flagged to developers. It will go into our database of scenarios and test cases we track. We’ll have more information about this on the desktops, but from what I saw on the screen it looks like we detected [the cones] correctly, but for some reason the planner was conservative and decided not to change lanes. We’ll create a scenario that says, here, the right thing would have been to change lanes, and the next versions will have it addressed.”
A few minutes later we turn left across five lanes of oncoming traffic onto California Street and reach our destination: an open-air market called the Milk Pail. Rather than stop, though, we continue back toward the Google campus. At one point Dolgov and Torcellini realized air wasn’t coming out of the A/C system because the vents weren’t on. That was the biggest problem the car encountered until we’d just about reached campus. I had about closed out hope for more excitement when Dolgov makes an announcement.
“We wanted to make the ride a little more interesting for you,” he says.
• • • • •
Dmitri Dolgov is soft-spoken with (at least on the day we met) biblical patience for a reporter’s repetitive questions. He arrived at Google in 2009 at the same time as Chris Urmson. They’d known each other from their DARPA challenge days, then as adversaries. Dolgov was part of Sebastian Thrun’s group at Stanford. Evidently the rivalry still lingers; when I met everyone later that day to discuss my ride, they brought it up unprompted.
“I was on a team that was not Chris’s,” says Dolgov.
“Came in second,” says Urmson.
“Different years, different places.”
“Same year, different places.”
“Well,” says Dolgov. “At least we didn’t flip our car upside down.”
Race history aside, they share a clear belief that the self-driving car will have a transformative impact on road safety. Dolgov has been quoted as saying that if the car has to fail, he hopes it will “fail gracefully.” When I ask him to elaborate, he brings up the incident with the road work cones.
“It didn’t handle it as well as you would want to,” he says. “But it kind of failed gracefully. It saw the cones early, it slowed down smoothly.” One could imagine a less graceful car, say, plowing right through them. “The car needs to recognize its limitations and do the conservative thing given its limitations,” he says. “Even when that means being slower or being stuck.”
The Google car is programmed to be the prototype defensive driver on city streets. It won’t go above the speed limit and avoids driving in a blind spot if possible. It gives a wide berth to trucks and construction zones by shifting in its lane, a process called “nudging.” It’s extremely cautious crossing double yellows and won’t cross railroad tracks until the car ahead clears them. It hesitates for a moment after a light turns green, because studies have shown that red-light runners tend to strike just after the signal changes. It turns very slowly in general, accounting for everything in the area, and won’t turn right on red at all—at least for now. Many of the car’s capabilities remain locked in test mode before they’re brought out live.
“We have lots of things we turn off until we’re confident,” says Dolgov. “And if you had a self-driving car that handled everything else well but didn’t do right on red? That’s still a useful thing.”
Google’s self-driving “drivers” are programmed for caution, too. Torcellini, who’s been behind the wheel since 2009, may have logged more driverless miles than anyone on the planet. He has a breezy manner—in the Google car movie he’ll be played by Paul Rudd—but the driver training program he’s designing is a rigorous one. He recruits detail-oriented and disciplined individuals, several with military backgrounds. (“You can’t have a Craigslist ad for people with that type of experience,” he says.) He screens them with a driving interview. Once hired, drivers go through at least a month of training in both classroom and car, and must pass regular performance tests to ensure a steady development.
We’re hugging the curve when suddenly we jam on the brakes—a utility truck has cut us off on the left.
“It seemed like I had the easiest job in the world, just sitting around in a Lexus, but in fact we’re paying really close attention to what the system is doing,” he says. “We know we have the reputation of not only Google but also the technology [on the line] every time we take a car out of the garage.”
This safety-first culture gets a big assist from Google’s developers, who don’t need the cars to leave the garage to put them through several types of off-road simulation. They can invent a world using their CarCraft system to test out any road scenarios imaginable. They can tweak the code and model hundreds of thousands of miles to determine what effect a change would have over time. They can even take an instance when the driver disengaged and see what would have happened if the car had been left alone.
Inside the car, I found out what that means in practical terms: Google drivers don’t have to get into an accident to learn from one.
• • • • •
When Dolgov said they’d made the ride “a little more interesting” for me, he meant the team had staged a series of scenarios to demonstrate the full scope of the car’s city street capabilities. First we turned down a road and came upon a woman riding a red, green, and yellow Google bicycle in the shoulder. She held out her left arm, which the car’s windshield camera detected, which the software then identified as a turn signal. A little yield sign appeared above the cyclist on Dolgov’s laptop, and the car slowed down until the cyclist cut left and out of harm’s way.
The car then passed a few more staged tests. We slowed for a group of jaywalkers and a rogue car turning in front of us from out of nowhere. We stopped at a construction worker holding a temporary STOP sign and proceeded when he flipped it to SLOW—proof the car can read and respond to dynamic surroundings, making it less reliant on pre-programmed maps. We merged away from a lane blocked by cones not unlike the one that stumped us earlier.
Urmson cites three big technological advances that have facilitated the car’s shift to surface streets. The first is its ability to classify the objects around it. Early on, he says, they would be lucky to distinguish a car from a pedestrian; now they can not only tell the difference but determine their travel paths. The second (and related) improvement has been in machine vision. That helps the car not only react to signals it expects, such as traffic lights, but those it doesn’t, such as the STOP/SLOW sign. The third step forward is in machine learning—the system’s ability to interpret data and resolve a problem on its own.
One of the clearest examples of its progress is the way the car turns left. Andrew Chatham, the mapping lead, explains that two years ago, the car made all left turns the same way: it drew a fixed path through the intersection and adjusted its speed accordingly. But over time the team realized that cars approaching a left turn at a green light follow a very different path than those starting from a stopped position. So now the computer recognizes this situation and computes a new route on the fly. It’s those little tweaks that bridge the gap between a jerky robotic ride and an amazingly smooth one.
Toward the end of my test run, after about a half hour of uneventful city driving, the car enters a cul-de-sac at the end of Charleston Street. We’re hugging the curve when suddenly we jam on the brakes—a utility truck has cut us off on the left. A few moments later it becomes clear that Torcellini had disengaged auto mode and hit the brakes manually; the car probably had another second to decide on its own whether or not to stop, but rather than take the chance it wouldn’t, Torcellini performed what he calls a “conservative takeover.” I certainly hadn’t seen the truck coming, and the palpable release of tension in the car suggested this wasn’t one of the staged events.
“It’s very easy for us to go back and simulate what the car would have done, had we not disengaged,” says Dolgov, logging the incident. Later on I ask Torcellini what he thought would have happened if he hadn’t taken over, and instead left the car to its own devices. “I think it would have stopped,” he says. “It would have done the exact same thing I did.”
• • • • •
Urmson met us after the ride to see how it went. I said I knew I wasn’t supposed to compliment the car, but that the ride had felt amazingly smooth. He turned to Dolgov.
“Oh,” he says, “you told him the first rule of self-driving.”
Urmson seemed a little disappointed that we’d needed to take manual control of the car twice. He says it took about six months of focusing on surface streets to get the basic foundation in place, but that accounting for all the nuances of city driving will take more time. “Driving where you did today, it’s unusual that we would have disengaged twice,” he says. “Compared to some of the situations you’ll see on the road, a lot of what you saw today was pretty benign. It’s stuff in your daily life, you might drive it without worrying about it too much. So now we’ve still got room to grow there, but we’re pushing again on a few more of these longer problems. Trying to deal with smaller streets, less room to maneuver, more difficult intersections—that kind of thing.”
It’s still too soon to declare victory in the race for driverless cars, but that hasn’t stopped some experts from saying they expect autonomous vehicles on the road by 2030 (Nissan has pushed up its timeline to 2020). The history of self-driving technology is filled with premature confidence. At the 1939 World’s Fair, the famed General Motors’ Futurama exhibit predicted a world of radio-guided cars by 1960. In his recent New Yorker story on the Google car, Burkhard Bilger writes that one of the team’s lead engineers, Anthony Levandowski, keeps reminders “of all the failed schemes and fizzled technologies of the past.”
Urmson knows all too well the hurdles that still remain. One of the main limiting factors is that any city where the self-driving car can go must first be mapped with a precision far greater than what even Google Maps achieves. That’s doable in Urmson’s mind: “We know how to deal with that scale of data,” he says, referring to Maps and Street View. A greater challenge may be processing and codifying the myriad subtle social cues that remain so vital to navigating crowded city streets. Right now the car can’t detect a driver trying to wave it into a lane, for instance, or someone requesting a merge through eye contact. And it still can’t understand that universal language of urban traffic: honking. (It is, however, developing an “ear” for sirens.)
Then there is the matter of scale. Google has a goal of roaming all of Mountain View in the self-driving car by the end of this summer. That would be no small feat: the city has the feel of a typical college town, which makes it a great launching point for many midsized U.S. cities, and its population of 74,000 no doubt rises considerably during the daytime hours, when the car roams its streets. But no one is mistaking it for San Francisco or New York or any major metro area where traffic is so tightly packed and street behavior so wildly unpredictable that a super defensive driver might suffer from paralysis by indecision.
“There’ll be lots of little wins between here and there, but that’s the big one.”
Still, Google is keenly aware what’s at stake. There’s the safety component, with cities recognizing the need to strive for zero traffic fatalities. The nature of urban mobility itself is also on the line. Larry Burns, a former vice president for research and design at G.M. who’s now a paid Google consultant, says taxi-like fleets of shared autonomous vehicles can become viable business models if they can capture just 10 percent of all city trips. “I think that should be viewed as a new form of public transportation,” he says. Having recently invested in the ride-sharing service Uber, Google no doubt senses that marrying urban travel demand with autonomous vehicles could transform car-ownership as we know it.
I asked Urmson when he’ll consider the car a success. “I think it’s a success when people are using it in their daily lives,” he says. “When we have cars out there and people are moving around and we have statistical data that says we’re saving more lives than had these people been driving themselves. The first time somebody who doesn’t work for Google is riding in one of these cars, getting to Grandma’s house or to work in the morning, or moving when they couldn’t otherwise move around the city, that’ll be a huge day for us. There’ll be lots of little wins between here and there, but that’s the big one.”
A few days later I got an email from the Google press staff saying the self-driving car team had run a computer model on the near-miss with the utility truck. Turns out the car would have stopped on its own with “room to spare.” That sounds like one of those “little wins” Urmson mentioned, but I doubt if he celebrated much. There’s a rule about that, and besides, the car already knew.
All images courtesy Google.