From Darpa Grand Challenge 2004Clash of the Headless Humvees

But Levandowski's master-of-the-universe savoir faire is undercut by his need to prove himself -- to himself, to his demanding father, to the world -- by taking on the supersize project. Consequently he has a tendency, hardly unknown in the robotics world, to overpromise and underdeliver. And that's how I found myself on my first visit to Berkeley at a grassy testing facility outside of town watching him put his motorcycle through its paces -- except it wasn't the real thing, the Yamaha 125cc dirtbike that was sitting in his workshop, but a foot-long one-fifth-scale model operated by remote control. "We've actually tested this thing on campus," Levandowski explains. "We've had it run into professors by accident, which was amusing, but if it had been the real thing, we might have killed someone."




The little cycle actually provides a great demonstration of the physics that Levandowski is up against. The bike doesn't want to stay upright, so when you give it a shot of juice from the remote, it tends to go into a wild wobble and then dramatically crash to one side or the other. Unless you know what you're doing. The trick, as any motorcyclist will tell you, is proper countersteering. If the bike is falling to the right,
you turn the front wheel to the right, more sharply than if you were making a right turn, and that provides the counterforce to bring the bike back to vertical. Turn too forcefully to the right, as I did when I was trying to drive the mini-motorcycle, and you'll overshoot vertical and wipe out on the left side or get locked into a desperate, doomed left-right oscillation that results in a similar slide across the ground.




Finally, though, I got the hang of it -- and sent the miniature bike flying out of sight over a ditch, a thrill straight out of Thelma & Louise.






Time is running out. There are, Red Whittaker reminds his Monday-morning Robotics 101 class (CMU's academic cover for the Red Team), only 109 days left to race day. Whittaker turns his attention to the student responsible for designing how the decals of the various corporate sponsors are going to fit on the car. That may sound trivial, but Whittaker calls it a "hot-button issue." Despite its reputation as the rich kid on the DARPA block, the Red Team's aggressive massage of corporations for money and equipment is the only way it can afford to expend, all told, the equivalent of more than $2 million to win a $1 million prize. "I'm working on it," the decal specialist replies, which even a first-time visitor can sense is not the right answer. "Could you not get 'working on it' but get it done?" Whittaker replies sharply. "It would be a tremendous 'done.'"




If only the other vehicle issues could be solved with a neat wrist slap. Of all the problems that Whittaker and his team face, perhaps the most intractable is one that has bedeviled artificial intelligence for years: how to get the damn thing to see. By the mid-'80s, roboticists at CMU and elsewhere had worked out the dialogue between a mounted video camera and some basic computer steering algorithms that allowed cars to be driven without drivers. Basic pattern recognition, being able to recognize and follow the straight line of a road, is something that computers tend to be good at. But having a clue about what's going on in the dynamic, three-dimensional and completely unpredictable off-road environment is something else again. Dodging a boulder in the middle of the path shouldn't be too hard. But recognizing whether the large object is in fact a boulder or a clump of tumbleweed or a moving Grand Challenge competitor will be. Ditto distinguishing the edge of a desert path from the rough country that lies next to it, or recognizing "negative spaces" like ditches or potholes.




The Red Team has chosen a range-finding laser, or LIDAR, to be the vehicle's primary "eye." The LIDAR shoots out a pulse of light, which hits something and bounces back, providing the onboard computer with a good idea about the size and shape of the world outside. Twin video cameras (for stereoscopic, near-3-D vision) and radar fill out the picture, the radar working on the same bounce-back principle as the LIDAR but with less accuracy.




In fact, every serious Grand Challenge contender has come up with a sensing system that looks on paper very similar to the Red Team's. Whittaker's contention is that it's not about the hardware; it's how you use it. The team is working on software to intelligently "fuse" the sensory data -- for instance, when the vehicle hits the inevitable cloud of sand, it needs to know to stop listening to the LIDAR, which gets hopelessly flummoxed by airborne particulates, and start relying on the radar, a champ in the dust. An enervating number of man-hours have been spent on what in layman's terms might be called a steady-cam system. All the sensors are being mounted on gimbals that keep them taking pictures at a fixed, precisely calibrated angle to the ground, no matter whether the Humvee has its nose in the dirt or the sky. The onboard computer can then construct a precise and reliable 3-D image of the world from the rapid-fire barrage of images it receives. That's the theory anyway. The reality is that the more the sensors bobble around, the likelier the vehicle is to run itself into a ditch.




Of course, it helps that the Humvee, with its racing suspension and 10 inches of suspension clearance, doesn't roll so easily. The route-finding systems don't have to be perfect as long as the beast can steamroller the small stuff.

Everybody loves an underdog, especially if the "underdog" is a group of affluent, well-connected 15-year-olds -- and so the tech corporations of southern California have opened their checkbooks and their inventory catalogs for the Road Warriors. American Honda has been the chief sugar daddy, courtesy of PV parent Tom Laymon, providing the team with an Acura MDX whose excellent computerized handling controls and navigation system are proof that the gap between the car industry of today and the autonomous military vehicle of tomorrow is less than might at first be supposed.




The SUV's inaugural desert test begins inauspiciously as we overshoot the turnoff on Highway 15 east of Barstow that will take us into the Mojave Preserve. This is an off-road vehicle, Robertson says: Just go over the shoulder, pull a U-ey, and head cross-country back to the exit. I'm driving. We cautiously plow into the dirt shoulder and nose down an incline, a large rock slamming into the undercarriage at the bottom. "No problem, we just scraped the exhaust pipe a bit," Robertson says, a blithe spirit if there ever was one.




Among the Road Warriors' strengths is home-field advantage. Palos Verdes is just a three-hour drive from Barstow, and a couple of the adult mentors have been able to recognize exact course locations in some of the desert photos posted on the DARPA Web site, DARPA's effort to scramble the backgrounds notwithstanding. So we visit a few.