Stories of a golden future pepper science fiction – where computers make life better and menial robots serve you martinis as you return home from a long day at work. Unfortunately, given the state of the art in computer intelligence, the robot would be unlikely to find its way from the bar to the dining room. On your return home, you would be greeted by the sound of breaking glass as the robot yet again misses the doorway. The world can be a horribly confusing place for a young robot. My research teaches them to see and make sense of the world, which includes doorways.
Robots have a rather narrow view of the world. To see it through a robot’s eyes, imagine taking a photograph of your kitchen and examining it through a microscope. What you had always thought looked like a doorway is suddenly a collection of red, green and blue dots, barely distinguishable from the dots making up the rest of the room. The problem is that, like a robot, you are only seeing a small part of the picture at a time. Teaching a robot to see the bigger picture has proved surprisingly difficult.
Rest assured that there are more noble motives here than a love of fresh and unspilt martinis. A robot with the capacity to navigate a house bearing a martini could navigate the surface of Mars bearing a mining probe. Some friends have suggested that after navigating my house, Mars may not be so difficult. Without the intelligence to recognise what they see, the robots used last year in NASA’s Mars Rover mission were little more than highly accessorised remote-controlled cars. The robots of the future will need to be far more canny.
So how do you teach a robot to make sense of the world? Let us return to the photograph of your kitchen and look at it again though our imaginary microscope. As the microscope moves around the photograph you will find yourself instinctively grouping bits of the image together. First you find a clump of dots which are a similar shade of green, then a region which seems to have a particular texture. The human mind has a marvellous talent for finding patterns like these and identifying parts of the photograph that belong together. From this we then work out which objects these patterns correspond to: a person, a fridge, a doorway.
For a robot to make sense of the picture, it needs to be able to find patterns too, and recognise these patterns as objects. Identifying the parts of the photograph that belong together can be rephrased as a question of mathematics. Translating human instinct into instructions that a robot can follow then becomes a manageable question of probabilities and statistics. Identifying which objects these patterns correspond to, however, is as yet unsolved. At present, my robot will happily spend the entire day offering the fridge another martini.
Like many graduate students, I am met with blank stares when trying to discuss my research at parties. Most scientists suffer this because their work seems impenetrably complicated. I lie at the other extreme. The problems that my robots solve seem trivial even to a child. But they are incredibly rewarding. Each time that the robot recognises something new, it feels as though I have been given a passing glimpse into how human minds solve these problems. My friends have occasionally found me standing in the middle of a party, staring at the entrance to a room and mumbling, “But how do I know that it is a doorway?” I explain this as the frenzy of research. My friends blame the martinis.
The winning essay in the Wellcome Trust/New Scientist 1998 Millennial Essay Competition for postgraduate students.