Posted 05.27.2011 at 11:23 am 3 Comments
PR2, Playing
Robots have particular trouble with abstract concepts, and even with connecting the dots between two very similar concepts. As such, they have trouble interacting with objects that they haven’t specifically interacted with before. So even if a ‘bot knows how to manipulate, say, a soup spoon, a larger salad spoon might confuse it even though the two are--at least in concept and in usage--the same thing.
There are two ways to overcome this: teach a robot to become familiar with every possible object in the world, or create algorithms that help a robot figure out how to manipulate certain objects on its own, without prior experience. The latter option is naturally ideal, and that’s what the PR2 Playpen is all about.
A conveyor belt keeps a stream of objects coming into the robot’s workspace, where the algorithms go to work helping PR2 study and assess them. PR2 then tries to manipulate them. Autonomously, the robot tries different strategies to pick up each object and then to return it to the workspace. As you can see in the video, it’s a hit-and-miss approach, but with each object more data is collected that can be used to improve the manipulation algorithms. (This should come in handy should Willow Garage respond to PopSci’s plea for a PR2 for our offices, as we have lots of cool toys for it to play with.)
Bonus: the Playpen requires no adult supervision, so PR2 can play all day and all night, building on its algorithms around the clock. Think about that as you’re laying down to sleep tonight: while you are resting comfortably, somewhere out there the machines are getting smarter. Sweet dreams.
[IEEE Spectrum]
"The AI does not hate you, nor does it love you, but you are made out of atoms which it can use for something else." ~Yudkowsky
Shameless Screen Grab including priceless comments from Popular Science