SAN JOSE, Calif. – Stanley, the robotic car from Stanford University that triumphed in a recent $2 million race across desert terrain, learned to drive in much the same way as any 16-year-old: by following the lessons of experienced humans.
It worked. The Volkswagen Touareg R5 crossed the finish line ahead of four other competitors in the DARPA Grand Challenge race for driverless vehicles, sponsored by the Pentagon, on Oct. 8.
When the Stanford team first started testing Stanley, a blue sport-utility vehicle, he had a 12 percent blunder rate for “false positives” – incorrectly assuming 12 percent of the objects in front of him were obstacles big enough he had to swerve around them.
So the team instructed Stanley’s software to take notes while a human driver maneuvered the car over different types of terrain. By following this guidance, the false positive rate dropped to one in 50,000 objects.
The race was on
This kind of debugging, conducted during 1,200 miles of off-road testing in the deserts of Southern California and Arizona, put Stanley first across the finish line in Primm, Nev., after traversing a 132-mile course with no human intervention.
The victory was particularly sweet because Stanford beat heavily favored Carnegie Mellon University, which entered two vehicles that came in second and third.
The race was funded by DARPA, formerly the Defense Advanced Research Projects Agency, because the Pentagon wants to make one-third of the military’s land vehicles self-driving within a decade.
But the really big market for robotic driving systems is in the civilian world, for both safety and convenience.
DARPA ran the first Grand Challenge race in March 2004 with disappointing results; no vehicle got farther than eight miles. I covered that race and was disenchanted enough to pass on attending this year’s repeat. So much for my powers of prediction.
Instead, I caught up with the Stanford team leaders in early October during a break at a robotics conference held in a hotel near Fisherman’s Wharf in San Francisco.
I discovered that the Stanford team, which didn’t enter last year’s race, still picked up pointers on what not to do.
First, DARPA had said the off-road course would be navigable by a standard off-road vehicle such as a pick-up truck or SUV. Some teams wasted time and money designing custom vehicles.
Second, the Stanford team believed standard off-the-shelf hardware would be sufficient for the sensors and computers. Again, some teams got caught up with custom designs.
Third, the Stanford team realized the race was really all about software. Completing the course would require a vehicle that could accurately understand its position, what was coming on the road ahead and how to respond.
“We focused on artificial intelligence from the get-go,” said Stanford team leader Sebastian Thrun, director of Stanford’s artificial intelligence lab.
How they did it
Stanford started with an off-the-shelf vehicle, the Touareg donated by Volkswagen’s Palo Alto research lab. The Palo Alto lab also provided engineering support for connecting Stanley’s computers to the steering wheel, gear shift, gas pedal and brakes.
The only physical modifications to the vehicle were minor: adding a skid plate to protect the undercarriage, a beefed-up front bumper and heavy-duty tires.
Stanley constantly updated its position and orientation by taking readings from three global positioning system receivers, as well as from motion sensors, and by measuring the number of times its wheels rotated.
DARPA provided a rough outline of the 132-mile route by giving each team a computer file containing 3,000 way points just two hours before the race started – an average of 23 way points per mile. But simply following this electronic bread crumb trail wasn’t enough to win, because the way points weren’t always centered on the road and didn’t show how to get around obstacles.
To navigate between way points, Stanley had a vision system consisting of five laser range finders and one video camera. The lasers looked at the ground immediately in front of the vehicle, while the camera looked further down the road. The motion sensors helped the vision system by telling when the car was pitching up and down, allowing the vision system to correct for the movement.
The position and visual data was fed into a planning module that examined every possible path and picked the best solution – somewhat in the same way chess software evaluates every possible move on the board.
The planning module also included a “road finder” feature that looked at the road immediately under the vehicle and determined which section of the forward horizon looked the same. If Stanley was running on a brown dirt road with green bushes on the sides, it would aim for the stretch ahead that was brown rather than green.
Finally, a controller module turned the selected route into specific instructions for steering the vehicle and setting its speed.
Stanley’s programmers also had to make lots of small but critical decisions in setting rules for the road. One example: Stanley was told to drive over rocks no higher than six inches, but to avoid rocks any larger.
Intel donated six computers with Pentium M chips to tackle all the processing this required, and all six computers were installed in Stanley’s trunk. But the team needed only two computers to do the actual work, and one more computer to keep a log. The other three computers just went along for the ride.
The $2 million prize goes to Stanford, rather than any of the 60 individuals who worked on the team, and Stanford will also own any patents that come out of Stanley.
New meaning to valet parking’
Thrun, 38, believes robotic driving will someday advance to the point where human geography is changed. No longer will big office buildings or shopping centers need to be surrounded by big parking lots. Instead, cars will dutifully park themselves on the edge of cities, waiting for a call to come retrieve their owners.
Long before then, robotic systems will save thousands of lives by taking the wheel at the last moment to prevent accidents caused by driver error.
Thrun confidently concluded: “It’s a no-brainer to me that cars will drive themselves.”
Comments are no longer available on this story