Cornell University robot Kodiak reads your need for a beer

PUBLISHED : Saturday, 01 June, 2013, 12:00am
UPDATED : Saturday, 01 June, 2013, 5:35am

Computer scientists at America's Cornell University have created a robot that can tell if you want a beer and pour it for you. The same robot can also guess whether students are hankering for a cup of coffee instead and pour it for them.

Kodiak the robot was as handy with a latte as with a lager, could open refrigerator and microwave doors, and even tidy up, say the researchers. In tests, the hard-wired humanoid correctly anticipated a student's next move between 57 per cent and 82 per cent of the time, depending on how far into the future it was "anticipating".

Programmers broke down human movements, uses of objects, possible trajectories of motion and the intentions of activities, such as eating dinner or having a beer, said Dr Ashutosh Saxena, an assistant professor who works on personal robotics.

Along with graduate student Hema Koppula, Saxena built an algorithm around 120 videos of 10 common activities, giving the robot the flexibility to "rate" the next move in a sequence and move accordingly.

"You can think about human activities as a document in which there is a basic alphabet of what people can do, and we sequence things together to do long-term activities," Saxena said. "We all do very basic things, like move our arms, get up, hold something, eat something or drink something. Now we can put this together in a sequence to do a variety of things."

Equipped with a 3-D camera, the robot would visualise the student's motions and the location of key objects, such as a coffee cup, Saxena said. The tasks were long, and some were complex, such as eating an entire dinner. Kodiak had to learn to pour beverages without interfering with other activities, or assist someone trying to microwave a dinner.

The robot performed best when predicting an action within a second. But because the possibilities multiply, accuracy dropped to 71 per cent in the three-second time frame and 57 per cent in a 10-second time frame, according to the researchers, who will present their work at the International Conference on Machine Learning in Atlanta.