The ability to acquire and then share knowledge is a central component of human culture and civilization. A small milestone in the exchange of robot knowledge has now been demonstrated by two bots working in different academic research labs.
Researchers at Cornell University previously devised an online game, called TellMeDave, through which volunteers can help train a robot to perform a task and associate different actions with commands given in everyday language. By guiding the robot through a task, a volunteer trains a machine-learning algorithm so the robot can perform the task again. And this learned behavior is stored in a central repository called RoboBrain that’s accessible by other robots (see “The World’s First Knowledge Engine for Robots”).
Some time ago, through this platform, a type of research robot called PR2 had been taught to perform a number of a simple demonstration tasks, including picking up several mugs from a table and placing them on top of upturned bowls. Several hundred miles away, in a lab at Brown University, a different type of robot, called Baxter, has taken what PR2 had learned and used it to figure out how to perform the same task in a different setting.
The work is part of an effort to figure out how robots might share information in useful ways. That could reduce the need for meticulous reprogramming, and it could allow robots to adapt quickly when faced with a new task or an unfamiliar setting.
“It’s pointing in an interesting direction,” says Stefanie Tellex, an assistant professor at Brown University, whose group enabled the Baxter robot to learn. “When you put a robot in a new situation—and in the real world it happens in every room the robot goes into—you somehow want that same robot to engage in autonomous behaviors.”
Speaking last week at the Bay Area Robotics Symposium, held at the University of California, Berkeley, Ashutosh Saxena, who led the development of TellMeDave and RoboBrain, said that robots will increasingly share information in the future. “We are trying to make robots learn and share knowledge,” he said. “Different robots can push and pull knowledge from the [RoboBrain] database.”
The key challenge in transferring learning between the robots at Cornell and Brown was that they are physically completely different, which means that low-level commands, such as those specifying the position each joint needs to assume in order to reach for a mug, will not match. Tellex’s group had to figure out a scheme that would allow commands to be transferred between the two platforms.
Ultimately, she says, it would be ideal for a robot to figure out how to translate information for itself, based on how its physical body compares with that of another robot. “This is what we’d all like to do, and this is really a baby step toward that vision,” Tellex says. “There are a lot of remaining technical challenges.”
Nick Roy, a professor at MIT’s CSAIL says many researchers, including several at MIT, are interested in enabling robots to share knowledge. One thing making it possible, he says, is increased bandwidth and cloud computing capacity: “As we’ve gotten the ability to handle more and more data across the Internet, it’s become more feasible to have this kind of shared [knowledge] representation. It’s something that the robots community has long aspired to.”