Go bot racer, go! - The Darpa Grand Challenge

Senior Editor

Navigate a race-prepped Humvee 250 miles across the desert from Barstow, Calif., to Las Vegas in 10 hr or less and collect $1 million in cash: Sounds reasonable, right?

Try it without the driver.

That's what Red Whittaker and his Red Team from Carnegie Mellon hope to do March 13 at the Darpa Grand Challenge. They aren't alone. The quirky race billed as "A Historic Demonstration of Autonomous Robotic Vehicles" has attracted some 40 teams, each with its own robot-piloted design. The entry quickest to complete the grueling desert course within the allotted time wins. Autonomous means exactly that: Race vehicles can receive no external communication or human control whatsoever. Just to make it more interesting, organizers will reveal the route just 2 hr before the start of the race.

Audacious? Perhaps. But "Dwelling on qualifying descriptions such as 'difficult' or 'unprecedented' doesn't get it done," says Whittaker. "People overdo the sensing/navigation aspect of the design. Robots have been navigating for two decades so there is plenty of information out there." Whittaker knows. It was his robots that scoped the fried nuclear reactor at Three Mile Island and more recently shimmied down abandoned coal mines.

What's probably less debatable is his choice of vehicle to make the trek: a 1986 military-issue Humvee. Humvee's off-road prowess is legend, and the battle-weary, high-mileage specimen that arrived at the University last spring is receiving a major overall. A larger-displacement, turbocharged powerplant will likely replace the original (and anemic) 6.2-liter diesel. Having more oomph helps offset imperfections in the computer control and coordination systems and, of course, may give it an edge over the competition.

Special attention was lavished on the vehicle's suspension. Here, Red Team took cues from off-road gurus Chad and Ron Hall at Team Hummer. "We traded-up to race-grade air suspension and springs to soften the ride," says Red Team Member Chris Urmson, a fourth-year engineering doctorial candidate at the University. "We're also running special B.F. Goodrich Baja tires and Hutchinson rims."

Whittaker explains all the fuss over ride quality:
"Human drivers on rough terrain do a great job of stabilizing their visual field. We take a three-step approach to stabilize perception sensors on the truck."

First, the specialized suspension improves the overall ride. Next, an intermediate mechanical stage that occupies what would be the front seats holds the navigation equipment, computers, and electronics. Isolation mounts from Lord Corp., Cary N.C., protect it and the enclosed equipment from shock loads, though the motion would still make you seasick. That's the reason for step three.

Mounted on top the intermediate stage is an actively controlled three-axis gimbal fitted with a fiber-optic gyro and harmonic servoactuators from HD Systems, Hauppauge, N.Y. The gimbal holds a steady inertial gaze for key perception sensors that attach to it. In essence, the intermediate stage provides a stable platform from which the gimbal can operate. Otherwise, the gimbal would oscillate and servo wildly to compensate for bumps and jars, which is unacceptable. "A jitter-free, rock-solid gimbal is one of our Big 10 winning edges," says Whittaker.

Steer this way

Several types of sensors will help the Red Team Humvee (code named Sandstorm) find its way to Las Vegas. Overall position information comes from a GPS receiver fitted with two disc-shaped antennas. A technique called carrier-phase differencing figures relative position between antennas and GPS satellites, which boosts vehicle position accuracy. Included in the package are an IMU (inertial measurement unit) and an encoder driven from the drivetrain half shaft. A controller combines all these signals to further boost position accuracy. The IMU also acts as a backup should the GPS signal blink out.

An airborne-class laser range finder looks 50 m ahead of the Humvee over a 60° field of view. A spinning reflector in the device fans out the laser light while a sensor measures time of flight for reflected beams to give distance. The distance data feeds to software that builds models of the terrain in real time and in tandem with stored digital USGS maps, satellite images, and other digital terrain data gathered by the Team. (It is here the route will be programmed in when it becomes known 2 hr before the race). A separate lidar (light detection and ranging) system handles close-in modeling and shores up the peripheral field of the laser scanner.

Next is millimeter-wave, continuous-frequency radar with a 200 to 350-m range. It takes over when the lasers become blinded by desert dust, though resolution is much coarser than with the lasers. Rounding out the perception-sensor package is a stereovision video system with a 20-m range for precision, up-close maneuvers.

Having different sensor types gives the multilayer navigation software some redundancy. "If the primary laser or stabilization platform fails, we'll use the lidar or radar, and finally, the stereovision cameras," explains Urmson. "Should all that go awry we'll run blind off GPS waypoints. In any case we may have to slow or cut our safety budget."

Running the C++ and C-based code requires plenty of computing horsepower. But current ruggedized, military-grade computers aren't capable of real-time mapping. The Team opted instead for a high-end Itanium server. Such computers normally sit in clean, air-conditioned rooms and don't get bounced around the desert. This underscores the importance of the shock-isolation and air-filtration systems. No air conditioning will be needed because ambient temperatures near Las Vegas average about 75°F in early March.

The Itanium is one of seven onboard computers linked by a gigabit Ethernet network. A single-board PC 104 format computer, for example, handles some of the low-level control such as steering and throttle, pointing of the laser, and an axis to stabilize the radar.

No cakewalk

Early last December the bright-red Humvee drove 4 miles on its own guided only by GPS signals, considered an important first step. Obviously, the chances for malfunctions rise as more perception sensors and components come into the mix.

One such component is race-logic strategy software from a gaming company. The software helps figure, among other things, vehicle position relative to the competition, when to pass and with what aggression, and if you are out front, how far to lead the race.

Urmson likens Red Team's overall race strategy to that of the tortoise in the fabled tortoise/hare race: slow, steady, and reliable. Average speed is expected to be about 25 mph with bursts to 35 mph.But even at such a modest pace there are still plenty of potential pitfalls. For example, a deep ditch directly behind a big hill would be impossible to see beforehand and the truck could fall in. "Having the robot understand that in some situations it must slow to see what's ahead is a complex process and a concern," says Urmson. "At the same time, it can't be too cautious so we maintain speed."

Whittaker has his own set of concerns: "Technical ambition is probably one of the biggest things that could take us down. It isn't about how much technology you can load on. At some point it's a convergent monotonic focus on what it takes to win, and having the courage to cut -- painful as it may be -- things that don't."

And then there are the skeptics who say no entry will finish the race. "DaVinci probably heard similar things about his inventions," Whittaker scoffs. "You know all there is to it? Drive from Barstow to Vegas and get the million bucks."

Make contact