Robot learns by playing and imagines its own future

BERKELEY, Calif. (KGO) -- A robot at UC Berkeley is doing something that isn't robotic at all: it's learning about the world by playing with toys and then imagining its own future.

"The goal of this work was to enable a robot to learn entirely on its own," said Chelsea Finn, a Ph.D. candidate in computer science.

Finn and fellow Ph.D. student Frederik Ebert wrote the machine learning code that allows the robot named Vestri to figure out the laws of physics by playing with objects all by itself.

"Similar to how a baby can learn how to play with toys and learn how toys respond when the baby picks it up, pushes it, that sort of thing," Finn said.

Vestri spent a week moving things around on a wooden table to gather the initial "experience" for the first phase of research. It's a very different learning model from a robot we saw in the same lab two years ago, putting various-shaped blocks into matching holes and getting better over time. In that model, humans "graded" the robot on how well it did at the task.

"Kind of like how you would train a dog how to do a trick," Finn said. "You would give them feedback, say this is good, this is bad, but that takes a lot of effort in terms of the human const supervising the robot and telling it what to do."

Vestri needs no human input to learn, and actually imagines the outcome of its actions in blurry simulated video images before picking the action most likely to succeed.

"You can see that the robot is imagining pushing this car to the front and here you can also see the exact motion that it predicted," Ebert said as the robot moved its gripper next to a toy car, then presented three blurry pictures of what might happen next.

Because of the way it learns, the robot can even interact with objects it's never seen before. We asked it to move our somewhat fragile ABC7 News microphone across the table and hoped for the best. After some grinding of gears, Vestri showed us images of a three possible futures -- including two that appeared to depict the robot's gripper breaking our microphone in half.

Ebert explained to us that the robot was really just trying to decide if the microphone, with a thin black handle connected to a large blue ABC7 logo flag, was one object or two. It correctly decided the microphone was a single object, and moved it to the goal on the first try. If the object slips or gets stuck, the robot patiently stops, imagines its future again, and tries a different approach.

Ebert said the real world potential is huge: "for industry, for elderly care, households and so on."

Unlike a factory, where robots already thrive, homes and hospitals are environments that constantly change, with objects moving around and new ones showing up.

"It needs to have a lot more than just the ability to do the same thing over and over again," said UC Berkeley computer science professor Sergey Levine. "It needs to have a kind of a common sense."

This robot already knows moving one object will also push the objects behind it. Next, Finn and Ebert want to teach it to pick things up, lifting them over the stuff that's in the way.

"But this is definitely future work," Ebert said.

The concept of a robot that learns like a baby is, after all, still in its infancy.

Click here for more stories, photos, and video on robots.