X

Cheers! PR2 robot knows where to pour your beer

You know summer's around the corner when robots are helping you grab a cold one. Now, an anticipatory algorithm is helping them refill your glass when you're ready for another round.

Tim Hornyak
Crave freelancer Tim Hornyak is the author of "Loving the Machine: The Art and Science of Japanese Robots." He has been writing about Japanese culture and technology for a decade. E-mail Tim.
Tim Hornyak
2 min read
Bottoms up: Willow Garage's PR2 shows how humanoid robots can be indispensable. Screenshot by Tim Hornyak/CNET

Robot! Fetch me a beer!

Yes, robots can actually carry out that order. Now, they can even anticipate where to pour your beverage of choice.

Cornell University's Personal Robotics Lab has trained a PR2 robot from Willow Garage to figure out where and when to pour beer, as well as perform other actions that require anticipation.

Armed with a Kinect 3D camera and a database of 3D videos, PR2 can analyze what it sees by breaking down activities into several steps. Then it anticipates what might happen next with objects it picks out in the scene. It can choose the most likely next step for activities like eating, drinking, cleaning, and putting things away.

For instance, the video below shows the robot anticipating what a man carrying a pot near a fridge wants to do and moving to open the door for him.

Without using the anticipatory algorithm, the video shows PR2 spilling beer on a table. But when the algorithm is working, it's able to anticipate where a cup will be and pour the beer into it.

PR2's ability to anticipate actions correctly degrades with time: it made correct predictions 82 percent of the time when looking one second into the future, 71 percent correct for three seconds, and 57 percent correct for 10 seconds.

"Even though humans are predictable, they are only predictable part of the time," Cornell Computer Science Professor Ashutosh Saxena was quoted as saying in a release. Saxena plans to discuss the research at robotics conferences in the U.S. and Germany next month.

"The future would be to figure out how the robot plans its action. Right now we are almost hard-coding the responses, but there should be a way for the robot to learn how to respond."