Robots Go to War

What a difference a year of development made. Many of the participating teams returned with vastly improved systems that made the best of the learning experience that the first race really was. During prerace testing for the second Darpa Grand Challenge, the TerraMax hightailed it at 35 mph through the scrub near Barstow, California, with its steering wheel magically moving back and forth as if it were being manipulated by the Invisible Man. During the actual race, staged in October, the truck was one of five vehicles to complete the 131-mile desert course. The $2-million prize was claimed by an autonomous Volkswagen Touareg SUV, called Stanley in honor of its alma mater, Stanford University."Yes, we were competing against each other," says Sebastian Thrun, leader of the triumphant Stanford Racing Team (and a 2005 Popular Science Brilliant 10 winner)."But all of us had to defeat nature. The real victory is that five vehicles finished. I think this will prove to be a pivotal moment in the history of transportation."

Stanley averaged 19 mph over the race and hit a top speed of 38 mph. It was equipped with five ladar sensors that reached an average of 85 feet and a single camera with a range of 260 feet. The biggest challenge the â€bot faced during the race was a flock of birds that confused it by temporarily occupying a lakebed it was crossing. Nevertheless, Thrun is convinced that a version could be readied for military-convoy duty almost immediately and that civilian highway applications are just a matter of time."If nothing else," he says,"the Grand Challenge showed people that it's possible."

Ladar and stereo cameras proved the most effective sensors for the Grand Challenge entrants; the jury is out on which is more promising. Other UGVs feature systems that detect color, heat and texture. Eventually, the goal is to be able to differentiate not only between wispy and solid, and living and inanimate, but between friend and foe."If it's a child, you want to stop," says Donald Verhoff, Oshkosh's executive vice president of technology."If it's a guy with an RPG-7 [grenade launcher], you want to run him over."

But better sensors create more data to process, which increases the time it takes to make decisions, which forces the 'bot to slow down. This places a premium on software that"stovepipes" directly to the correct solution rather than working through an endless series of possibilities. The most practical way to do this is to code up algorithms that establish a rules-based hierarchy covering every situation that might arise. Obviously, this requires a boatload of
if-this-then-that commands. And even so, UGVs are bound to encounter circumstances that weren't anticipated by the programmers.

At Darpa, Larry Jackel, who manages three of the agency's ground- robotics programs, is spearheading a project to program UGVs with behavioral cues that will allow them to respond appropriately to unexpected situations and learn from experience."We believe the problem is too complex to write an algorithm to traverse every piece of topography," Jackel says."A robot's got to be like a child, who learns to crawl, then walk, and then run." But this approach remains highly experimental, and it seems unlikely to be found on the next generation of military UGVs.