Move Over, Herbie

"This is the first endurance race in history where the machine will have to make all the decisions," said Sebastian Thrun, associate professor of computer science and leader of the Stanford Racing Team.

The Defense Advanced Research Projects Agency (DARPA) launched the Grand Challenge in 2004 to speed the development of autonomous off-road vehicles. Making robotic vehicles that move quickly over rough terrain is an open challenge, and one the military needs to solve to reach its goal of having one-third of vehicles driverless by 2015.

Car manufacturers also are eyeing the competition with interest. "Driver assist" functions increasingly use computer controls in anti-lock brakes, adaptive cruise control and electronic parking aids; obstacle-avoidance and lane-keeping technologies may soon be added to that list.

"We're in this competition as a community that tries to gain knowledge, that tries to make roads safer," Thrun said.

The race will follow a course revealed just two hours before the start. Abandoning their creators at the starting line, entries ranging from all-terrain vehicles to modified golf carts will roll out across the desert.

The first robot to complete the course in less than 10 hours wins $2 million.

False start

The first DARPA Grand Challenge in 2004 drew 106 entries; teams ranged from major research institutions to dedicated backyard hobbyists. Agency Director Anthony Tether said he hadn't seen this level of excitement in national-security research since the days of the Apollo space program. But the race itself was a washout.

Frontrunner Carnegie Mellon University's modified Humvee went almost 8 miles before driving too close to a cliff's edge. The vehicle's wheels spun helplessly until a tire caught fire and officials in a chase vehicle turned the machine off. A team of engineers from San Diego entered a modified Jeep, but its global positioning system (GPS) malfunctioned on race day and the vehicle circled aimlessly. Palos Verde High School in Southern California fielded the only high school team; its entry, Doom Buggy, lived up to its name by crashing into a concrete barrier soon after the starting line.

This year's competition drew 195 hopefuls, almost twice as many as last year.

"The field is a lot more serious this year, and the general belief is that someone is going to make it," Thrun said.

Thrun, hailing from Carnegie Mellon, joined the Stanford faculty in 2003; he now directs the Artificial Intelligence Lab. Thrun was between jobs when the first race was announced, he said, but last July he approached the School of Engineering to ask about putting together a team.

In fall 2004 the Computer Science Department offered a course, DARPA Grand Challenge, in which 30 students helped build a prototype robotic vehicle.

The team has snowballed to 50 members, including eight faculty and 17 external members. Stanford partnered with Volkswagen of America's Electronics Research Laboratory in Palo Alto, which donated the car and modified it for drive-by-wire control. Local venture capital firm Mohr Davidow Ventures heads up organization and communication. Other sponsors are Android, Intel, Honeywell and TYZX.

First HAL, now Stanley

In a field that includes military vehicles and bizarre Frankentrucks, Stanford's robotic entry, nicknamed Stanley, is a sophisticated family car that wouldn't look out of place at a Palo Alto shopping mall. It's a gray Volkswagen Touareg, a state-of-the-art sport utility vehicle modified with oversized wheels, a reinforced bumper and a protective metal plate under the car.

The fuel-efficient turbodiesel engine allows the vehicle to complete a 10-hour race and power all its electronics on a single tank of gas. And unlike many competitors, Stanley is fully street legal.

"That was an important design constraint to facilitate the testing of the robot," said vehicle lead Cedric Dupont of Volkswagen. To do trial runs, team members pile into the car, drive to Barstow, Calif. (an off-roading mecca outside Los Angeles), and set Stanley loose.

So far, the Stanford team has made almost a dozen weekend road trips. After a day out testing in the desert, members usually congregate at the Slash-X--an off-road biker bar where they have gotten to know the staff--for late-night research sessions.

"We'll basically take over a corner of the bar, open up our laptops and start programming," said Gabe Hoffman, a doctoral candidate in the Aeronautics and Astronautics Department who is designing control software.

The only information DARPA has released about this year's course is that it will start in Southern California. For Stanley's maiden voyage in December, the vehicle drove the first 8.5 miles of the 2004 course. At one point the robot slowly navigated a tricky curve, bumping down a rocky dirt road with a sharp drop-off to one side, while team members watched from a distance. Although the team didn't put the Touareg through the entire 2004 racecourse, this early display of autonomous ability showed Stanley to be a contender.

Advertisements for Volkswagen cars use the slogan: "Drivers wanted." The sign on Stanley's side panel reads: "Drivers not required."

Where the rubber meets the sand

The strategy for autonomous driving is deceptively simple. A robot must know where it is, see the road ahead, choose the best course of action and act.

An array of sensors plus GPS determines Stanley's position within 2 inches. Lasers mounted on the roof continuously scan the ground in front of the vehicle looking for cattle gates, ditches, barbed-wire fences or disabled cars (possibly competitors) on the course.

But lasers let Stanley see only four car lengths ahead. At racing speeds up to 35 mph the vehicle uses long-range radar to scan the horizon. Stereo-vision cameras, mounted inside the car to avoid dirt spattering, take high-resolution pictures of the course, while computer algorithms "learn" the terrain and map out an optimal driving surface.

"This is really a software competition," said Thrun.

Monitors inside the vehicle show the robot's frenzied thought process as 10 times a second it analyzes and weighs the relative safety of many possible routes. A probabilistic algorithm developed by postdoctoral researcher and software lead Michael Montemerlo interprets the flood of incoming sensor data.

Stanley's brains--seven Pentium M laptop computers--are in the trunk. The machines are shock-mounted to survive the bumps on the road, and the networked system has multiple copies of every program. A thick spinal column of electrical wires connects the computers to an electronic brake, throttle and chain-driven steering column.

"If one computer fails, or if even two computers fail, we'll still be able to continue in the race," explained David Stavens, a doctoral candidate in computer science and member of the software group.

The road ahead

On May 10, the team met DARPA officials in Barstow for a site visit, during which Stanley successfully navigated a trash-can-littered course.

This qualifying round will narrow the field to 40 teams. Another qualifier in September will further winnow the contestants to 20 finalists. DARPA officials are not allowed to comment on the visits until June 1, but Stanford team members returning last week were overjoyed. The robot's performance, they said, was "flawless."

In preparation for race day, the team plans to log 2,000 autonomous miles. Over the next four months team members are likely to be found making adjustments to Stanley in the garage, debugging code in the lab or heading down to the desert for test runs.

Stanford's superior software experience and big-name sponsors are matched only by the team's "infinite enthusiasm," Thrun said.

"We're in it for winning," Thrun said. "No less than that."