It's the day before the DARPA 2005 Grand Challenge, and Sebastian Thrun, team leader of the Stanford team, is having a blast. He's cheerfully conducting a one-man orchestra of media attention, race preparation, and technology evangelizing. As he escorts me into the pit area--a section of the Buffalo Bill Resort and Casino parking lot where all the teams are camped out in tents next to their autonomous vehicles--he tells me that this challenge is most certainly, absolutely not about winning.
Thrun, who is the director of the Stanford Artificial Intelligence Lab, firmly believes that "if anyone finishes the race, everyone here has won." He tells me that his goal in building autonomous vehicles has little, if anything, to do with their possible military applications. He believes that self-driven vehicles are the way to create safer roads for everyone, and the way he talks about the development of Stanley, the team's souped-up Volkswagen Touareg, it's clear that this is an exercise in the most joyful form of geekery. Stanford didn't compete in the first DARPA challenge; Thrun says they actually chose to watch the race rather than compete, so they could spend as much time as possible building the perfect robotic entrant.
DARPA Grand Challenge entrant Stanley
The car itself is equipped with four laser range finders, a radar system, two stereo cameras, and a monocular vision system, all of which allow it to build a 3D image of the path through which it travels and help it detect turns and obstacles as it chooses the correct path. A blue box mounted in the back of the VW controls the drive functions and, like all the vehicles in the race, Stanley is equipped with an E-stop required by DARPA, which lets them shut down the bot at any time or pause it to allow another vehicle to pass.
The Stanford vehicle runs on seven Pentium M computers, all running Linux, mounted in the back of the Touareg, although Thrun tells me they discovered they really only need two computers to run the vehicle--Stanley is just that efficient. The vehicle's computers are extremely low power, too, drawing all the power they need from the Touareg's engine, eliminating the need for a generator on the vehicle. Cedric Dupont, a senior research engineer at Volkswagen, tells me Stanley is probably the most environmentally friendly bot on the course--it runs on biodiesel, and its good fuel efficiency means the team doesn't have to put in an additional fuel tank. In fact, aside from the computers in the back, the bank of radar and electronic "eyes" on top, and the red button in the center console that switches from robotic to manual drivetrain, the Touareg is very nearly stock. However, Dupont confides that the Stanford team has actually gone through six of the roughly $40,000 vehicles in the course of the project--something Volkswagen wasn't really expecting when it got involved. Three are in attendance at the race--Stanley, a backup replica called Stanlette, and a jet-black chase vehicle.
As a dramatic side note, Thrun is a former Associate Professor of Computer Science and Robotics at Carnegie Mellon University, which is camped out in the pit right next door. Carnegie Mellon is the brainpower behind Red Team and Red Team Too, whose red Humvees are the heavy favorites to not only complete the course, but take home the $2 million. In contrast to the Stanford team, which numbers 60 including, Dupont says, "everyone," the two Red Teams comprise 250 people. While Stanford is camped out in a tent, preprogramming Stanley on an IBM ThinkPad, the Red Team is working out of a temperature-controlled trailer in the middle of the parking lot, which is filled with computers. They've invested, according to rumors in the pit, some $5 million in the two Humvees, and their sponsors include such heavyweights as Google and Caterpillar. DARPA regulations require Carnegie Mellon to enter as two teams if it's going to run two vehicles, but grad student Josh Johnston tells me the two Humvees were built in the "same development pipeline."
DARPA Grand Challenge entrant H1ghlander
Carnegie Mellon has been here before. Its 1986 military Humvee, Sandstorm, raced in the first DARPA challenge and was the most successful bot on the course--it completed 7.5 miles of the 142-mile course. This year, Carnegie Mellon is reentering Sandstorm, along with a 1999, consumer-model H1 named, appropriately, H1ghlander. Johnston says the team chose a Humvee for the first DARPA challenge because of the vehicle's ruggedness, but says its military history didn't hurt, in a military-sponsored competition.
H1ghlander also features seven Pentium M computers (also running Linux), and both vehicles house stereo cameras, laser-range scanners, and radar equipment in large, carbon-fiber domes mounted on the roofs. It's clear that while Stanford's Thrun presents a cheery, geeky, utopian vision of autonomous vehicle development, Carnegie Mellon is all business. It's serious about the competition and likes its chances.
While Stanley had four perfect runs in the qualifying events that helped place the bots for the starting line, Sandstorm and H1ghlander turned in faster average times. H1ghlander will be the first bot out of the starting gate, followed by Stanley, and Sandstorm will leave third. By Friday, the day before the race, Stanford has clearly become the emotional favorite for spectators and media alike.
But if Stanley tugs at the heartstrings, Team TerraMax captures the imagination with sheer bad-ass-ness.
DARPA Grand Challenge entrant TerraMax
This bot in question is a 32,000-pound MTVR--a medium tactical vehicle replacement. Team leader Gary Schmiedel tells me that Oshkosh truck already makes this vehicle for the Marines. The truck is big enough that all the electronics are mounted in the cab and on the front of the truck, leaving its bed available for hauling 7.1 tons of cargo offroad or 15 tons onroad. The truck stands 9 feet tall, and its three monster-tire axles boast independent steering, so it stands with its wheels cockeyed to demonstrate its steering potential.
Schmiedel tells me it's as maneuverable as a Humvee, but TerraMax's size has already posed some problems. For one thing, it's slow. Although the truck can book through a straightaway, qualifying rounds featuring narrow entryways or tunnels mean TerraMax has had to creep through at a snail's pace, or even back up and reenter several times. Secondly, although TerraMax technically finished the qualifying rounds in fifth place, DARPA has pushed the bot back in starting position to at least 19 out of 23. The problem? The truck is so big that when it's in front of the other bots, they view it as an obstacle, throwing them off their game entirely. DARPA will start the vehicle later in the race so that the other bots won't become convinced there's a gigantic wall dead ahead. In sum, this truck is awesome.
The pit is filled with interesting stories, dune-buggy-like underdogs, and brilliant hopefuls, and even though they've had 18 months to prepare, most of the teams say they don't expect more than a handful of vehicles to complete the course the next day--whatever that course ends up being. They'll receive a CD--called an RFFD--that contains the course's waypoints and mapping data at 4:30 a.m. the day of the race, and the first bot will launch with first light, hopefully around 6:30 a.m.