Highlights from Wired.com:

A robot has taught itself to smile, frown, and make other human facial expressions using machine learning.

To get the incredibly realistic Einstein robot to make facial expressions, researchers used to have to program each of its 31 artificial muscles individually through trial and error. Now, computer scientists from the Machine Perception Laboratory at the University of California, San Diego have used machine learning to enable the robot to learn expressions on its own.

...

To begin teaching the robot, the researchers stuck Einstein in front of a mirror and instructed the robot to “body babble” by contorting its face into random positions. A video camera connected to facial recognition software gave the robot feedback: When it made a movement that resembled a “real” expression, it received a reward signal.

“It’s an iterative process,” said facial recognition expert Marian Bartlett, a co-author of the study. “It starts out completely random and then gets feedback. Next time the robot picks an expression, there’s a bias towards putting the motors in the right configuration.”

After the robot figured out the relationship between different muscle movements and known facial expressions, it started experimenting with new expressions, such as eyebrow narrowing.

The robot’s expressions are still a bit awkward, but the researchers say they’re working on ways to make them more realistic, as well experimenting with strategies besides “body babbling” that might speed up the learning process. The group says its studious robot may even improve our understanding of how infants and children learn to make facial expressions.

“The idea is to try to understand some of the computational principles behind learning,” Bartlett said. “Here the computational principle is reinforcement learning and active exploration, which may also be behind learning motor movements in an infant.”

The next step is to get the Einstein robot to start socializing. Once the robot can mimic facial expressions in a social context, the researchers plan to use him in an “automatic tutoring” experiment.

“We’re putting facial expressions onto the robot so that he can engage with a pupil in a non-verbal manner and approximate one-on-one human tutoring as much as possible,” Bartlett said. “Studies have shown that human one-on-one tutoring improves learning by as much as two standard deviations — we want to know how can you try to approximate that with robotic tutoring.”

Full Article

A video of the robot: