Walking Bots

Darpa (The Defense Advanced Research Projects Agency), the Pentagon's mad science division, wants to teach little mechanical puppies to think. Hopefully, that'll let the bots run around with soldiers on the battlefield one day. Getting robots to maneuver around rocks and trees and potholes is tough -- just ask any of the tinkerers whose bots bit the dust during last year's "Darpa Grand Challenge." That was the all-drone, off-road rally across the Mojave Desert, scheduled to last 150 miles. No robot made it past mile seven. One way to get the bots to go further, some drone-makers think, is to give their creations legs, so that they can maneuver just like a person or an animal would. But that's easier said than done. Walking, it turns out, requires a zillion tiny calculations to keep balance and avoid obstacles. It's so complex, Darpa notes, that "handcrafting the control laws and parameters" needed for robots to hike "may not even be possible with reasonable effort." So instead, Darpa would like to get the bots to figure out how to walk on their own. In the Learning Locomotion program, algorithms will be created that learn how to locomote based on the experience of a legged platform confronting extreme terrain. It is expected that the performance of these algorithms will far exceed the performance of handcrafted systems, creating a breakthrough in locomotion over extreme terrain. Further, it is expected that these algorithms will be broadly applicable to the class of "agile" ground vehicles.
Darpa is planning on handing out a series of $600-800,000 contracts to try to teach drones to walk. And the robots the agency wants researchers to train are 6.6 pound, 10.6 inch-long "Little Dogs." During the 15-month first phase of the "Learning Locomotion" project, Darpa wants the pooches to be able to travel .6 of an inch per second, and scale obstacles about 2.5 inches tall. For Phase II, those numbers should go up to approximately 3.8 inches and 5.7 inches, respectively. That may not sound like much. Bu the drones will have to be smart enough that that can "learn 'on-the-fly' how to traverse new obstacle types," Darpa tells researchers. "Government tests will measure the ability of the performer systems to learn from experience." "Learning Locomotion" is part of a series of Darpa efforts to come with computers and robots that can think for themselves. The agency is sinking $29 million into creating a "Perceptive Assistants that Learn" -- software-based secretaries that understand their bosses' habits and can carry out their wishes automatically. Lockheed got another $6.6 million from Darpa to develop algorithms that will "enable computers to leap ahead of traditional information-processing capabilities used to perform cognitive tasks, such as deduction, reasoning, and learning," according to a company press release. A third program-- "Learning Applied to Ground Robots," aimed at smartening-up wheeled drones -- is looking for interested researchers now. There's also more than one Defense Department project involving dog-like machines. Last year, the Army doled out $2.25 million to two robotics firms to prototype a big, mechanical pooch capable of carrying ammunition, food and supplies into battle. little mechanical puppy