Scientists are trying a new tactic to train robots: babying them.
A team based at the University of Washington is coding robot behavior to look a lot like kids playing, reports Kit Eaton for Quartz. Typically, Eaton writes, engineers and researchers code robots to move in specific ways, sometimes by physically moving the machine by hand and recording the movements for later replication.
But by programming them to explore and experiment like children, the researchers hope that robots will be able to teach themselves. Though babies pick up a lot from watching their parents, they aren’t just imitators, writes Adrienne LaFrance for The Atlantic. By the time a child reaches 18 months, they can usually infer a person's goal, even if that person fails at the task.
"The secret sauce of babies is that they are born immature with a great gift to learn flexibly from observation and imitation," psychologist Andrew Meltzoff, a collaborator on the project, tells The Atlantic. "Roboticists have a lot to learn from babies.”
The team hopes to develop robots that can learn new tasks from watching and interacting with people, rather than just coding the task directly into their algorithms. "[T]he way to do it is through demonstration—showing the robot how to clean your dishes, fold your clothes, or do household chores," says project leader Rajesh Rao in a press statement.
The research team focused on two tasks when developing their algorithm: have the robot "look" where a human is looking and imitate someone moving toys on a tabletop, which will teach it how its movements affect other objects. They reported their process last month in the journal PLOS One.
Impressively (and perhaps slightly frighteningly), the robots did learn. LaFrance reports that the robot even developed its own strategies during the toy play to achieve the same results as the person it watched.
The steps may seem small, but so too are a baby’s first toddles. The clumsy robot of today just heralds the intuitive bots of tomorrow.