His task one afternoon is to clear a debris field. After many agonizing moments, in a set of abrupt and jerky movements, he crouches and with painstaking precision manages to grasp a two-by-four board and then drop it to his right. At the rate he is moving, completing the chore might take days.
Atlas in this case is an imposing, six-foot-tall humanoid robot that evokes the bipedal “Star Wars” robot C-3PO. It stands in a cluttered robotics laboratory here at Worcester Polytechnic Institute where a team of students, engineers and software hackers are training the 330-pound bundle of sensors, computers, metal struts, joints and cables.
Seven teams are working with Atlas robots, manufactured by Boston Dynamics, a small military-funded research firm based in Waltham, Mass. Like the others, the Worcester team is preparing for a December contest held by the Pentagon’s Defense Advanced Research Projects Agency. The contest is meant to accelerate work in the field of robotics by prototyping machines that can work effectively and autonomously in extreme emergencies, like the failure of a nuclear power plant.
The vision evokes decades of sci-fi movies like “I, Robot” in which self-directed walking machines glide through the world with grace and precision. At the moment the gap between that dream and reality is daunting.
The immensity of the challenge is underscored by the fact that, here in the lab, Atlas remains tethered — “on belay,” in the mountain climbing sense. Like a toddler learning to walk, it wears a safety harness, and whenever it moves, its human operators, equipped with safety glasses, position themselves behind a transparent plastic enclosure.
The care is warranted. If the robot did keel over, it would not be like a human football player hitting the turf. It would be more like a car crash.
And with each Atlas valued at roughly $2 million, the roboticists are going to elaborate lengths to make sure that doesn’t happen.
While there are smaller and lighter autonomous robots like Honda’s Asimo, or the South Korean Hubo, that do move around unaided, it is nearly certain that Atlas will never wander freely among humans. Each of its arms weighs dozens of pounds and feels like a metal baseball bat.
And when Atlas works, it howls. A hydraulic compressor screams at a decibel level just below Occupational Safety and Health Administration safety limits.
“It scares me,” said Matt DeDonato, the 26-year-old engineering leader of the W.P.I. team. “I don’t get close to it when it is operating.”
One member of the team holds the belay rope, and another stands guard by a large red emergency stop button, while the robot controller sits in front of a large computer screen. It displays a 3-D image, called a “point cloud,” that is generated by the robot’s laser range finder, which measures distance and maps objects.
With a keyboard and a mouse, the operator can issue commands to the robot, from walking to more incremental movements like opening and closing the three fingers of the machine’s hand.
To temper expectations for Atlas, the Defense Advanced Research Projects Agency is describing the first event to be held in December at Homestead-Miami Speedway as a “preliminary” trial. Teams will compete on a set of separate tasks like driving, closing a valve, crossing a rubble field and climbing a staircase. A year later they will compete at more complex tasks for a $2 million prize.
The teams will have the option of either controlling their robots manually or programming some of their actions to operate autonomously. Teams that choose to control their robots entirely by hand will be penalized, however, because the contest judges will intermittently limit the amount of wireless network bandwidth, occasionally blurring the controller’s view of what the robot is up to.
With just two months until the first competition, the teams are racing to program the robots just to complete the tasks, much less do them quickly.
“We’ve only had two or three months to play with our toys,” said Chris Atkeson, a professor in the Robotics Institute and Human-Computer Interaction Institute at Carnegie Mellon, a partner of W.P.I. “A graduate student usually has five years to do these things.”
The teams’ success will be determined largely by how well they program autonomous skills in the robots in the coming months. Mr. DeDonato, for example, said he was envious of another campus team, which in cooperation with Drexel University is programming a Hubo robot. That team has a sophisticated algorithm designed by the W.P.I. roboticist Dmitry Berenson, making it possible for the robot to use both arms simultaneously to close a valve.
“I wish we had that code,” Mr. DeDonato said.
For its part, Atlas would probably shrug — if only it could. But it can’t. Suspended in the lab and awaiting its controller’s digital commands, it is also missing the shoulder joint necessary for that motion.