Watch Darpa’s Headless Robotic Mule Respond to Voice Commands


If robots are ever really going to carry the bags of U.S. soldiers and Marines, the Pentagon’s futurists think, they’re going to have to act more like pack animals. That means responding to voice commands, figuring out when it makes sense to follow a human being and when it doesn’t, and getting around uneven terrain and other obstacles. Darpa thinks it’s off to a good start.

Over the past two weeks, the Pentagon’s blue-sky researchers took an upgraded LS3 robot — a cousin of the headless BigDog and PETMAN made by Boston Dynamics — on a walk through the woods and hills of Fort Pickett, Virginia. (LS3 = “Legged Squad Support System,” get it?) It was the first test conducted on the LS3 after bolstering the autonomy functions for the $54 million project, now in its fifth year. As the video above shows, when a human instructor calls out “engine on” and “follow tight,” the robot’s engine activates, and it follows its master on exactly the path the human takes. When the human calls out “follow corridor,” the robot will “generate the path that’s most efficient for itself,” explains Army Lt. Col. Joseph Hitt, Darpa’s LS3 program manager.
Darpa figures that it’s illogical to make a soldier hand over her rucksack to a robotic beast of burden if she’s then got to be preoccupied with “joysticks and computer screens” to guide it forward. “That adds to the cognitive burden of the soldier,” Hitt explains. “We need to make sure that the robot also is smart, like a trained animal.” So the LS3 uses a laser range-finder, specialized cameras and stereo vision to keep track of its human master.
Darpa is really literal when it means that the LS3 has to operate like a mule. It needs to haul up to 400 pounds, walk 20 miles and operate for 24 hours without human intervention. The program’s been undergoing tests for years, and Darpa and Boston Dynamics have put out numerous videos showing the gangly metal beast climbing woodlands, recovering from being kicked over, and running on treadmills. The Fort Pickett tests show much of the same. Only this time, the robot is way quieter — emitting “less noise than any combustion-engine system out there that’s tactically relevant,” Hill says — and its rounded back and spindly legs allow it to climb out of the uneven terrain that troops in, say, Afghanistan encounter. The LS3 can now also put its feet on uneven surfaces like logs, whereas before it labored to avoid them.
All this is the result of major advances to the LS3′s software, particularly as it integrates with the system’s sensors. At Fort Pickett, it walked through an obstacle course of shipping containers that subbed for the narrow alleyways of urban conditions. That showed the LS3 computing a course on its own: “The robot has to decide, ‘Can I fit here 20, 30 meters from now, or do I have to turn around?’” Hitt says. (Skip to 1:40 on the video to see what the robot sees.)
So far, the system responds to 10 basic voice commands — again, like a trained animal — like “stop,” “follow tight,” or “engine off.” But “perception and platform robustness” remain challenges, Hitt concedes, like ensuring the robot can react to changes in light or weather. It can’t, for instance, look at a snow-covered hill and figure out if the snow is too deep to traverse. Nor can it avoid battlefield dangers like gunfire or bombs. “We’ll have to always continue to add additional logic,” Hitt says.
That’s what’s going to happen over the next two years. Every three months, Darpa will take the LS3 to a different set of climate conditions that U.S. troops have to encounter; first up is the arid, desert environment of the Marines’ Twentynine Palms base in California’s Mojave Desert. That’s an indicator of Darpa’s desire to hand over its first robot to a Marine company in 2014. The LS3 may not have a head, but Darpa expects it to be at least as smart as a real-life mule.