The perception problem
It's not collecting the images, but making sense of them that counts when it comes to getting robots to recognise objects around it. James Mitchell Crow reports.
Seeing is much more than sensors. Our eyes gather light and our brains make sense of these signals. It’s the same with robots – it’s the smarts behind the sensors that count. And that’s where the step-change has taken place that allows field robots to roam the outdoors.
“What we always used to do in robotics was to try to build models of things,” says Hugh Durrant-Whyte, who now oversees research into machine vision and machine learning as part of his role as head of NICTA, Australia’s information and communications technology research centre. Researchers would try to build a computer model that captured the essence of a chair, for example, so that the robot would recognise a chair whenever it saw one. The problem is, even an object as simple as a chair is extremely hard to definitively describe. “Some have one leg, some have four, others have three, some just don’t look like chairs at all.”
The breakthrough came by sidestepping the problem and developing machine learning algorithms that give the robot the ability to build its own models.
So to teach the robot what a tree is, say, researchers now show it many pictures of trees, and let the robot pick out what is common about all these pictures.
Little humans learn about the world in the same way. “My description of a tree is actually every recollection I’ve ever had of looking at something I’ve called a tree,” says Durrant-Whyte. In the same way, the robot can learn to recognise buildings, cars, people, and all kinds of other objects.
The field is making progress. This year, for the first time, computer scientists Chaochao Lu and Xiaoou Tang from the Chinese University of Hong Kong published details of a computer algorithm that was better than people at recognising individual’s faces.
With computer memory now so cheap it’s just about possible for a robot to file away everything it ever sees, which suggest the robots of the future will be like us: life-long learners.