'Stanley' gets ready for the robo-desert race

But at Stanford University, you can find one going as fast as 40 mph on a dirt road in a sleepy part of the campus, and it goes by the name "Stanley."

"In the future, cars will drive themselves, no question," said Sebastian Thrun, a German-born computer scientist who is director of Stanford's artificial-intelligence laboratory and the man responsible for Stanley. As Thrun, 38, talked about the car last week, he was like a giddy schoolkid, particularly when he was hitching a ride with his robot car.

"The vision is to make cars safer, to benefit society," Thrun said, "but it's also to do something no one has done before."

Call Thrun and his Stanford Racing Team wild-eyed dreamers, but they're turning science fiction into fact. Or at least they're on the road.

Stanford will compete in the second annual DARPA (Defense Advanced Research Projects Agency) Grand Challenge, a U.S. military-sponsored desert race that tests the endurance of robots. It's a challenge that taps the creativity, brainpower and funds of U.S. universities and private-sector visionaries as the military tries to push the envelope of artificial-intelligence technology.

Last year, Carnegie Mellon University's robot Hummer, Sandstorm, went the furthest and fastest. But that's not saying much: It traveled 7.3 miles in a 144-mile race before burning out. No one claimed the $1 million prize. This year's race, scheduled for Oct. 8 and involving a 175-mile course, will net the winner $2 million, and the entrants sound like a major upgrade from last year. Sandstorm recently drove 200 miles over seven hours autonomously on a racecourse--a milestone for Carnegie Mellon.

On Saturday, the Stanford team raced Stanley over the 2004 DARPA Grand Challenge course and the robot drove autonomously for all but three miles of the 144-mile course in the Mojave desert. "Overall, we believe we achieved a major milestone," Thrun said on Monday. "The team is extremely pleased (and tired)."

This is Stanford's first time in the contest, and the 60-member team means business. Led by Thrun, a protege of CMU's team leader and robotics professor Red Whittaker, the Stanford team includes six core software engineers, mostly Ph.D. students. It also taps a group in Volkswagen's electronics research lab based in Silicon Valley. Members of Mohr Davidow, a prominent venture capital firm, have also invested in Stanley.

The workshop where Stanford's Racing Team tests and tinkers with its robot looks more than a little like a MASH unit. High-tech toys are piled everywhere. There's antique auto paraphernalia, old computer monitors, heaps of wires, scraps of metal and recycling bins overflowing with soda cans.

Stanley is actually a modified steel gray Volkswagen Taureg R5 that's only sold in Europe. It was donated by the German auto manufacturer. Other than the sensors that cover its roof, a control panel wedged next to the driver's seat and a computer network in the back, it looks like a regular SUV.

Except, of course, it doesn't need a driver. It's no easy task to teach a car to drive itself. It takes radar, vision and laser sensors fastened to the car that can act as early warning systems, detecting close and far-range obstacles. It draws on GPS (Global Positioning System) sensors to trace many of its steps.

The computer scientists have written more than 100,000 lines of code to tell Stanley what to do. There's a map that tells the car where to drive, a planning tool that points out unsafe terrain, and a controller that translates all that into action. The software runs on six Pentium M processors, which are Intel-made, low-power chips originally designed for the telecommunications industry.

"Humans are really good at recognizing obstacles," said Thrun. "Robots find this incredibly hard, like avoiding a rat on the road."

Any one of the computer functions can slip, causing the car to dodge a butterfly but then run into a rock.

During Saturday's trip, Stanley encountered six failures on the 144-mile terrain. According to Thrun, the failures were caused by false interpretations of the road ahead; an inability to drive through a couple of long underpasses with an extended GPS outage; a hardware failure of Stanley's cooling fans (the outside temperature hit 123 degrees Fahrenheit); and at one point, an inexplicable veering off the road.

And those errors were a step up from a recent trip. Two weeks ago, the team attempted the same course and Stanley made it only an average of 25 miles before stumbling, forcing a human driver who was along for the ride to take control.

Still, the team has made major breakthroughs in Stanley's software this month, according to Mike Montemerlo, a postdoctoral scholar in Stanford's AI lab who heads up software development for Stanley. New mapping code dramatically reduces false positives in the car's obstacle warning system. And the software now uses machine-learning technology to program the car to "remember" how it's first driven over a course manually, and then emulate those actions for autonomous driving.

Other new applications include commands for the car to slow down on rough roads. David Stavens, a computer science Ph.D. student, also wrote a program that tells the car to find the center of the road after it swerves to avoid an obstacle.

The recent road trip is one of about eight more the team will take before the October challenge, which will be held on a secret course somewhere in the desert.

"We will return to the desert soon with the hopes of fixing those remaining bugs, so that we might actually finish the course without any failures," said Thrun.

Other new applications include commands for the car to slow down on rough roads. David Stavens, a computer science Ph.D. student, also wrote a program that tells the car to find the center of the road after it swerves to avoid an obstacle.

Last weekend, the team headed back to the desert to drive over last year's course and test Stanley's endurance with its current improvements. The weekend road trip is one of about eight more the team will take before the October challenge, which will be held on a secret course somewhere in the desert.

"The vision is to make cars safer, to benefit society," said Thrun, "but it's also to do something no one has done before."

During last weekend's trip, Stanley encountered six failures on the 144-mile terrain. According to Thrun, the failures were caused by false interpretations of the road ahead; an inability to drive through a couple of long underpasses with an extended GPS outage; a hardware failure of Stanley's cooling fans (the outside temperature hit 123 degrees Fahrenheit); and at one point, an inexplicable veering off the road.

And those errors were a step up from a recent trip. Two weeks ago, the team attempted the same course and Stanley made it only an average of 25 miles before stumbling, forcing a human driver who was along for the ride to take control.

Still, the team has made major breakthroughs in Stanley's software this month, according to Mike Montemerlo, a postdoctoral scholar in Stanford's AI lab who heads up software development for Stanley. New mapping code dramatically reduces false positives in the car's obstacle warning system. And the software now uses machine-learning technology to program the car to "remember" how it's first driven over a course manually, and then emulate those actions for autonomous driving.

Other new applications include commands for the car to slow down on rough roads. David Stavens, a computer science Ph.D. student, also wrote a program that tells the car to find the center of the road after it swerves to avoid an obstacle.

The recent road trip is one of about eight more the team will take before the October challenge, which will be held on a secret course somewhere in the desert.

"We will return to the desert soon with the hopes of fixing those remaining bugs, so that we might actually finish the course without any failures," said Thrun.