PHOTO: Fumihide Tanaka/Machine Perception Laboratory/ University of California, San Diego

6 November 2007—Entertainment robots have become sophisticated enough that they can charm toddlers for weeks, or even months, and could soon be useful to teachers as permanent educational assistants, according to research reported this week in Proceedings of the National Academy of Sciences. Qrio, the dancing, bouncing, giggling robot spawn of Sony Corp., tried out its social skills on a group of children between 10 months and 24 months old at the Early Childhood Education Center at the University of California, San Diego, as part of a study on how children socialize with robots. The researchers found that the key to Qrio’s popularity was its ability to move and respond to the children in a way that was closely timed to the activity around it.

Robot designers usually measure the social competence of their creations by observing the robots’ performance in public demonstrations. Watching how well a robot can grab and hold a person’s attention gives engineers ideas about how to change software or aesthetic design. But getting good feedback from toddlers in public demonstrations can be hard, says Seema Patel, CEO of Interbots, a company that makes the entertainment robot Quasi. “It becomes trickier with toddlers, because often they’re being held or pushed in strollers,” and don’t have the freedom to interact, she says.

The research was done as part of the University of California, San Diego’s RUBI project, which has the goal of developing a robotic teacher’s assistant. Fumihide Tanaka, a researcher at the university put Qrio into the children’s play space and watched as the toddlers tried to sum up the new kid on the block.

The moment that a child began to engage with the robot seemed to depend very much on the robot’s responsiveness to social gestures. At first Qrio, whose actions were partially controlled by an unseen operator, would try to engage the children by waving a hand in front of them as they passed by, but usually the robot responded too slowly, and by the time it was waving, the kids had already moved on. Programmers soon scrapped that gesture. Instead they made Qrio giggle [mpeg video] immediately after the kids touched it on the head. The children quickly took more interest in the robot, suggesting that closely coupling the robot’s reaction time to the children’s behavior will enhance how they interact.

Especially with toddlers, who are just learning to speak and are still forming a concept of how a human looks and acts, appealing to emotions will be more engaging than trying to recreate human behaviors, says Patel. And this means making the robot react quickly with colors, sounds, and movements.

In another experiment, researchers tested the importance of random behavior by watching the children’s responses to the robot when it danced. Sometimes Qrio performed an elaborately choreographed dance, and other times it was allowed to jam to its own inspiration, which was actually a set of responses to optical cues coming into an internal camera. The kids, however, showed no preference for either the predictable, tethered Qrio or the jerkier, improvisational Qrio.

But predictability clearly factored heavily into how deeply the kids bonded with the robot, says Tanaka. Halfway through the experiments, the researchers reprogrammed Qrio to be much more predictable by limiting its set of behaviors. Even though they had been happily playing with Qrio for several weeks, the children stopped touching the robot as much and in general paid less attention to it until the programmers restored the full range of behaviors.

In designing a social robot, making it unpredictable may be even more important than making it seem lifelike, explains Patel. The easiest way to lose someone’s interest in a robot is to give it highly repetitive movements, she says. “The minute you start to see repeats, it’s clear that it’s not ‘real.’ ”

People also quickly lose interest in robots when their design guarantees a more elaborate performance than they can give, says Jonathan Klein, one of the developers of the BEAR (for Battlefield Extraction-Assist Robot), an army-contracted robot designed to carry wounded soldiers off the battlefield. Give a robot a mouth and you expect it to speak. Give it a vocabulary of words and you expect it to wield proper grammar and syntax as well. As soon as a robot takes on human traits, says Klein, people think it “is going to be as intelligent as you and me, have the same memory, the same capability, the same dexterity, the same subtleties of being able to pick up emotion.” Seeing a robot really dance is surprising and wonderful. But seeing it dance badly might be worse than nothing at all.

Making robots available for social interaction is a game of knowing your limits, and the best results are likely to come from picking a few modes of realistic interaction and making them as variable and unpredictable as possible, says Patel. Sony has stopped development of Qrio, but the study of the way it succeeds and fails to win over children’s hearts is informing others that are working on robotic teaching assistants. Kids will not only have to learn from these robots but adore them as well. And you can’t trick any 2-year old into doing that.

Video: National Academy of Sciences