Robot Learns Like a Child

Robot Learns Like a Child

NewsMax.com Wires
Friday, Jan. 26, 2001

EAST LANSING, Mich. (UPI) – Referring to their test prototype as their "baby," a team of researchers is teaching a robot to learn and think based on vision, touch and auditory signals.

As depicted in the article published today in Science, the device evokes the robot maid in "The Jetsons": a square box, with rounded "head" atop it, complete with video camera "eyes." This one's name is SAIL, short for Self-organizing Autonomous Incremental Learner.

What's unusual about SAIL is that it learns in the same fashion as humans, Juyang (John) Weng, one of the paper's authors, told United Press International. Robot-sitters – humans, of course – let the devices explore the world on their own. Certain behaviors are discouraged, as with a toddler, by pushing the "good" button or "bad" button.

In contrast, "past AI [artificial intelligence] guys envisioned a disembodied device residing in a conventional computer," explained Weng, a computer scientist at Michigan State in East Lansing.

An example of this earlier, knowledge-based approach would be the chess-playing computer famous a few years ago. Other early prototypes were spoon-fed human-edited sensory data while still being controlled by a task-specific program, an approach known as learning-based.

"These researchers thought a computer would not need eyes or a body," said Weng. "But gradually, we saw that to learn and grow, computers need to interact with the environment."

Although reluctant to be quoted by name until Weng's paper was officially released, one reviewer commented, "This is indeed a manifesto." Another said, "The argument for the paradigm is persuasive."

Weng terms his approach "autonomous development," meaning the device picks up information and incorporates it into future behavior as it goes along. Humans raise it, in Weng's vernacular.

"When a kid first opens its eyes, it doesn't understand what it is seeing. This comes later," Weng said. The same goes for his AI robot SAIL.

Developing a device that can learn autonomously requires designing a body appropriate for the intended working conditions – say, underwater. Then a developmental program is created, which starts up the device and could be considered its "birth."

At that point, the "baby" deviates from earlier learning-based systems in that real-time interaction with humans teaches it. For example, a researcher can push the device and it backs up. Or hand it a toy – it will take it and recognize it again.

"The aim is to enable robots to 'live' in the world," the researchers wrote for Science. "From infancy to adulthood."

Although this learning process takes time, Weng explains that it does not necessarily encompass 21 years as it would in humans, who are limited by biological constraints. The resulting devices are not task-specific and must learn their tasks from humans, rather than being programmed.

"Developmental robots," according to the researchers, "can 'live' with us and become smarter autonomously, under our human supervision."

What sort of humans will be "raising" these little servos? Weng, whose research is sponsored in part by the Defense Advanced Research Projects Agency of the Department of Defense, or DARPA, cites such potential uses as helpers to handicapped people, to perform space walks or tasks undersea, or to dispose of radioactive waste.

"First, though, they need a longer battery life," he noted.