April 29, 2015
To receive Newt’s weekly newsletters, click here.
On Monday I spent 20 minutes cruising the streets of Mountain View, California in a car that drove itself. Although the town was full of other vehicles, pedestrians, and bicyclists, neither human nor machine was hurt in the process.
The ride was a demonstration of Google’s self-driving car project, which for six years has been making steady progress toward teaching cars to drive without human intervention. It was exciting to see how close they are to achieving this goal.
Our host was Chris Urmson, who directs the self-driving car project at Google X, the company’s experimental research arm focused on developing breakthrough technologies to solve big problems. Chris and his colleague loaded us into the back seat of a Lexus SUV and instructed the car to choose a random destination in Mountain View. He pressed a button to begin and we were off.
The car drove itself significantly better than a new teenage driver, and better than a lot of experienced drivers, too. It traveled the speed limit, anticipated the moves of other drivers (often long before we would have noticed), and made complex judgements about whether it had time to turn safely across a lane of oncoming traffic or continue through a yellow light.
On a few occasions, the car applied the brakes in response to an unexpected swerve from a car in another lane or because of a branch hanging over the road. Chris explained that Google has designed its cars to be slightly more cautious than a human driver during the testing phase in order to avoid unnecessary risks. Often, however, the caution seemed more like the wisdom of a careful and experienced driver. On one occasion a large moving truck at an intersection made it impossible to see if there was oncoming traffic on the far side of the truck. The car nudged forward until it could peek around the truck to make sure the lane was clear before continuing ahead.
Google’s engineers have taught their cars to drive using machine learning, a technique that involves feeding all the sensor data from all of their vehicles into one central database so each car can make decisions based on the cumulative experience of the entire fleet. When one car encounters a new situation on the street that none of the others have seen before, the entire fleet “learns” from that new information almost immediately.
The result of several years of learning and improvement by Google engineers is a car that drives so much like you or me that it’s easy to forget the whole car is being controlled by a computer. It was a little surreal to ride along talking with Chris about public policy and his background as a computer science professor at Carnegie Mellon and then suddenly to remember, “Oh yeah, he’s not actually driving this thing.”
Chris explained that he began working on self-driving cars more than a decade ago with a group from Carnegie Mellon that came together to compete in the DARPA Grand Challenge. The Defense Advanced Research Projects Agency, an experimental wing of the Defense Department, established the competition for self-driving vehicles in part because so many of the casualties in Iraq and Afghanistan were from roadside bombs hitting supply convoys. DARPA held the first race for the early self-driving vehicles in the Mojave Desert in 2004. Chris’s team built a humvee that traveled the farthest–less than eight miles down the 150 mile course, before the truck got stuck on a rock and spun its wheels until the tires melted.
The teams made tremendous progress by the next year’s Grand Challenge, which was won by a team from Stanford led by Sebastian Thrun. After the next competition in 2007 (which required the vehicles to navigate an urban environment rather than the desert), Sebastian and Chris went to work at Google to accelerate the development of the technology.
As I was fortunate to see firsthand, the project has come a long way since then. Chris showed us prototypes of cars Google has designed from the beginning to be self-driving. They look a little like big Smart cars, but they’re also different from any car you’ve ever seen in a number of obvious respects. For one thing, they don’t have a built in steering wheel.
As I discussed in my book Breakout, fully self-driving cars will transform our lives and our cities in ways we are only beginning to understand. Clearly there will be enormous implications for traffic and safety if roads full of imperfect, easily-distracted human drivers are replaced with cars driven by computers that are always alert and always sober. And fewer of us may need to own our cars: you can imagine that one day cities will be filled with vehicles like Google’s prototypes, which you summon on your smartphone and direct to your destination much as you might call a taxi or an Uber today.
It was thrilling to see this glimpse of the future at Google and to spend time with some of the people who have pioneered it. I’ll be writing more Google X and some other amazing breakthroughs I saw in Silicon Valley in the next few weeks.
Sign up for Newt’s weekly newsletters here: