Robots Figuring Out how to Figure Things Out

You've heard it before: robots are bad at new things. They're bad at adapting to new situations, they're bad at recognizing new objects, and they're bad at coming up with their own ideas about how to carry out tasks in changing environments. One way to deal with this general ineptitude is to provide a robot with endlessly detailed instructions to minimize the amount of unfamiliar things it experiences. The world is an awfully big place, however, and if we want robots to be able to do robot stuff outside of tolerably controlled environments, such a brute force approach may not be the best way to go.

A much more elegant solution is to teach robots to think and adapt for themselves. Sounds easy, right? Right! I mean, wrong! It's not easy at all. But once you figure out how to do it, you can plop a robot down anywhere and ask it to do anything and it'll have a reasonable chance of figuring it out, or at the very least, be able to ask an intelligent question or two to get going.

At Cornell's Personal Robotics Laboratory, a research group is teaching a robot to generalize groups of objects, which is one of the most basic aspects of reliable adaptability. For example, instead of teaching a robot "this is a cup, and this is a slightly different cup" and so on, you can instead teach a robot to recognize features common to all cups, so that when it sees something cup-like, it can say to itself, "hey, that's a small container with a handle, I bet it's a cup!"

This same sort of learning method can also be applied to actions. By teaching a robot how to pick a few different types of cups, the robot can then generalize the lessons and apply them to completely new cups. And if you teach the robot to put a few dishes into a dish rack, it can then use what it knows about the objects and the rack to figure out how to put pretty much anything in there:

If you're the type who's impressed by numbers, this robot was able to put unseen objects into the right spot in the dish rack in the right way 92% of the time, which is about 92% better than I do. Touché dishwasher loading robot, touché.

[ Cornell ] via [ KurzweilAI ]



IEEE Spectrum’s award-winning robotics blog, featuring news, articles, and videos on robots, humanoids, drones, automation, artificial intelligence, and more.
Contact us:

Erico Guizzo
New York City
Senior Writer
Evan Ackerman
Washington, D.C.

Newsletter Sign Up

Sign up for the Automaton newsletter and get biweekly updates about robotics, automation, and AI, all delivered directly to your inbox.