The Age of Robots
We're close to making humanlike machines. It's time to reckon with the promises and perils
So how then to proceed? Increasingly, AI researchers are looking to children for the answer. Kids are essentially learning machines, and while no one is quite sure how they do it, there's clearly a lot of imitation and interaction, and plenty of room for trial and error. If robots are ever going to have humanlike intelligence, the new thinking goes, perhaps they'll have to develop it the way babies do. And that, says Cynthia Breazeal, the MIT roboticist behind Kismet's licorice-whip lips and Groucho Marx eyebrows, requires social interaction.
Kismet is programmed to seek sensory stimulation--voices, movement, brightness, and color--which it attracts with beguiling expressions and a sort of babbling baby talk. If an expression works, and a passing human comes up to play, Kismet's internal "social drive" is satiated. If not, the levels sink and Kismet tries a new strategy to connect. But there are balancing desires: Get too close and Kismet will let you know you've invaded its space with an exaggerated look of annoyance. Play too rough and the usually docile head may assume an alarming grimace or turn away.
"The whole point," says Breazeal, "is that the robot is trying to get you to interact with it in ways that can benefit its ability to learn." Basic movements are programmed into Kismet's behavior, but its handlers hope human feedback will help it learn new gestures and vocalizations by imitating people and storing successful attempts in its memory. "It helps the robot learn the social meaning of its actions," Breazeal says. The goal is for Kismet to learn not just to "think" for itself but also, as every child must, to understand that its actions have consequences.
What's good for the mind is good for the muscles. Maja Mataric, a computer scientist and roboticist at the University of Southern California, is trying to solve the problem of motor control. The entire range of human--or robot--mobility, she says, "can be collapsed down to a reasonably sized set of movements." Called "primitives," these basic motions can be combined or modified to produce novel activity. Take reaching. Whether you're stretching your racket arm out to volley a tennis ball or grabbing the lid off a boiling pot, says Mataric, "you use a standard way of reaching, the same basic movement." Once a robot has the basic moves down, mimicking people is much easier. All it takes is a little practice, plus a few learning and adaptation algorithms to help the machine capitalize on its mistakes. The result: a bot that can learn to dance the Macarena just by watching. In theory anyway--a good body is hard to come by, so Mataric works mostly with computer simulations with realistic physical properties such as gravity and mass.
Practical applications. Not all researchers buy the developmental theory of robot building. Kazuhiko Kawamura of Vanderbilt University, for example, programs his humanoid ISAC to perform practical tasks, such as feeding disabled patients. Machines that learn from the ground up might make for interesting interactions, Kawamura says, but "a humanoid robot needs to have more than just fascinating behaviors. To me, that approach only gets you a robotic baby. And we don't need a robotic baby."
advertisement
