A robot learns: Baxter’s video takes first place

我是walle

A robot learns: Baxter’s video takes first place

December 10, 2015 Contact: Kevin Stacey

Million Object Challenge

Factory robots that know the dimensions and location of an object can handle it, but grasping a variety of strange objects is a different matter. Can robots learn to handle strange objects through trial and error? Could they practice overnight, when they are alone in the lab?

For robots, picking up objects they've never encountered before is hard. Computer scientist Stefanie Tellex is working to teach them how to do it. A music video she made about the project won first prize in a competition and a new robot for her lab.
PROVIDENCE, R.I. [Brown University] — Brown University’s Humans to Robots Lab is about to get a new robot, thanks to Stefanie Tellex’s video-making skills.
Tellex, assistant professor of computer science and the lab’s principal investigator, entered the Rethink Robotics Video Challenge. The Boston-based company asked users of its “Baxter” robot to submit videos showing Baxter solving real-world problems, whether in research, manufacturing, or education.
Tellex took first place (with a little help from the band AC/DC and graduate student John Oberlin) for her video about teaching Baxter to manipulate objects from experience. The grand prize: A spanking new Baxter — the lab’s third — is on its way from Rethink.
On a factory floor, robots do a great job of picking up and manipulating objects that they’ve been programmed to handle. However, picking up objects that they’ve never encountered before can be a big problem for even the most sophisticated robots. Tellex has developed an algorithm that enables Baxter to learn how to pick up new objects by repeatedly trying (and often failing) to do so. Over time, the robot learns how best to pick up the object, and can do so successfully on future attempts.

本视频也可以通过百度云下载http://pan.baidu.com/s/1hriSfsg

Tellex is hoping to recruit other researchers who use Baxters users to scan objects as well. She’s calls it the Million Object Challenge.

“At night or on weekends, Baxters are just sitting there in labs doing nothing,” Tellex said. “If we could get all 300 research Baxters working on this during what would normally be down time, we could reach our goal of a million objects in only about 11 days.”
Tellex’s work has been the subject of recent stories on National Public Radio and in MIT Technology Review.