A wheeled robot released on a college campus has figured out how to open all kinds of doors and drawers while moving around in the real world.
The robot adapted to new challenges on its own, paving the way for machines capable of interacting independently with physical objects. “You want the robots to operate autonomously… without relying on humans to keep giving examples at test time for each new type of scenario you find yourself in,” says Deepak Pathak at Carnegie Mellon University (CMU) in Pennsylvania.
Pathak and his colleagues first trained the robot through imitation learning, providing visual examples of how to open objects such as doors, cabinets, drawers and refrigerators. They then set him loose on the CMU campus to try to open doors and cabinets he had never encountered before. This required the robot to adapt to each new object using artificial intelligence that rewarded it for figuring things out.
The robot typically spent 30 minutes to an hour learning to open each object consistently, says Haoyu Xiong at CMU, who built the robot and explored the campus for a wide variety of testing sites. The team included 12 training objects to practice with, then eight additional objects to test the robot's capabilities.
Even though its initial success rate was around 50 percent on average, the robot sometimes failed to open a new object on first startup. In the end, his success rate was around 95 percent.
In addition to learning on the fly, he had to be able to physically manipulate heavy doors, says Russell Mendonca at CMU. Achieving both goals costs $25,000, he says, which is much cheaper than other robotic systems with adaptive learning capabilities.
Robotic demonstration outside the lab 'marks a concrete step towards more general robotic manipulation systems,' says Yunzhu Li at the University of Illinois at Urbana-Champaign. “Opening doors and drawers – a seemingly simple task for humans – is actually surprisingly difficult for robots,” he says.