Defense dollars aid Carnegie Mellon's robotics research
A robot named Andy slowly scans a table and announces its goal in a distinctly British accent.
“I'm searching for a lug nut,” Andy says as Drew Bagnell and a team of researchers demonstrate their latest achievements in Carnegie Mellon University's National Robotics Engineering Center.
DARPA, the Defense Department's Advanced Research Projects Agency, is funding research on Robot ARM-S, short for Autonomous Robotic Manipulation Software Track, a system in which the robot can tackle tasks independently and learn from them.
The agency has funded a variety of Carnegie Mellon projects, including research on a flying car; development of an interactive, game-like curriculum for math and science education; and Mind's Eye, an ongoing project to perfect a surveillance system with intelligent software that will recognize human activities, flag unusual behavior and predict who might be dangerous.
In the ARM-S lab in a renovated 100-year-old factory in Lawrenceville, the lug nut exercise calls for Andy to “see” an object and pick it up in slow motions that mimic human movement.
“This robot is focused on perception integrated with manipulation. It's trying to see the lug nut and pick it up as opposed to executing a motion that does the same thing over and over again,” said Bagnell, an associate professor of robotics at Carnegie Mellon.
Eventually, the project calls for Andy to use lug nuts to mount a wheel on a frame. It can mount the wheel, but to secure it, it will need smaller hands.
When DARPA started the project in 2010, the agency said its goal was to develop robots that could perform military tasks without intense human oversight.
Intel Corp., the computer chip manufacturer, became a project sponsor when it saw the potential for development of domestic robots.
“Our goal for the next phase of this project is how we can apply the technology of this program to intelligent prosthetics,” Bagnell said.
Bagnell's team is working on the software that animates Andy, fashioning algorithms for the billions of calculations the robot must make for even the smallest task.
“It makes a lot of mistakes, but it is intelligent enough to see them, recognize them and do the next correct action,” explained ARM-S researcher Jean-Sebastian Valois.
“It does learn over time. That's machine learning. That's my field, to get robots to learn over time,” Bagnell said.