Robots get smarter by asking for help

To the right:Robot recharging itself

ASKING someone for help is second nature for humans, and now it could help robots overcome one of the thorniest problems in artificial intelligence.

That's the thinking behind a project at Willow Garage, a robotics company in Palo Alto, California. Researchers there are training a robot to ask humans to identify objects it doesn't recognise. If successful, it could be an important step in developing machines capable of operating with consistent autonomy.

Object recognition has long troubled AI researchers. While computers can be taught to recognise simple objects, such as pens or mugs, they often make mistakes when the lighting conditions or viewing angle change. This makes it difficult to create robots that can navigate safely around buildings and interact with objects, a problem Willow Garage encountered when building its Personal Robot 2 (PR2).

Where AI struggles, humans excel, finding this sort of recognition task almost effortless. So Alex Sorokin, a computer scientist at the University of Illinois at Urbana-Champaign, who collaborates with Willow Garage, decided to take advantage of this by building a system that allows PR2 to ask humans for help.

The system uses Amazon's Mechanical Turk, an online marketplace which pairs up workers with employers that have simple tasks they need completing. The robot takes a photo of the object it doesn't recognise and sends it to Mechanical Turk. Workers can then use Sorokin's software to draw an outline around an object in the image and attach a name to it, getting paid between 3 and 15 cents for each image they process.

In initial tests, the robot moved through Willow Garage's offices, sending images to be processed every few seconds. Labelled images started coming back a few minutes later. The accuracy rate was only 80 per cent, but Sorokin says this can be improved by paying other workers to verify that the responses are valid.

Sorokin believes his system will help robots learn about new environments. A cleaning robot, for example, could spend its first week in a new building taking pictures and having people label them, helping it to build up a model of the space and the objects it contained. If it got stuck, it could always ask for help again.

"This is a fantastic idea," says John Leonard, a roboticist at the Massachusetts Institute of Technology. Potentially this could allow robots to operate for long periods without direct intervention from a human operator, he adds.

The next step for the programmers is to enable PR2 to make sense of the human responses and then act upon them, Sorokin says.