Facebook’s robotic arms and legs are learning faster than ever

Robotics work can lead to improvements in other forms of machine learning, smartening Facebook's software for spotting bad online behavior.
(Loic Venance / AFP/Getty Images)

Although Facebook Inc. doesn’t sell robots, its researchers use plenty of them — and the company said its machines are getting a lot smarter a lot faster.

The social media giant announced Friday that its engineers, working with computer scientists from New York University, have reduced the time it takes to teach a robotic arm how to grasp objects to tens of tries, rather than hundreds or thousands.

It may seem like a tangential development for the social media giant, but advancements in robotics can lead to improvements in other forms of machine learning, smartening the software Facebook has begun to use to spot harmful or unfavorable behavior of users on the social network.

The company has been under increasing pressure to use artificial intelligence to police extremist violence, hate speech and misinformation on its platform. It has said that it is making progress but that systems that can reliably block such content without human intervention are still years away.

“The great thing about robotics is that it takes place in real time, in the real world,” Antoine Bordes, co-managing director of the company’s artificial intelligence research labs, said in an interview last week in Paris.


He contrasted this to research that taught AI to master games, such as chess or Go, which can be run at superhuman speeds, enabling a software agent to learn from playing millions of games against itself in a period of a few weeks. Many contemporary AI methods are extremely data-hungry, requiring thousands or millions of labeled examples to learn from — or thousands or millions of attempts in a simulated environment — to equal or exceed human performance.

Facebook began working on robots in the last year because it forced researchers to think creatively about how to make machine learning more efficient, Bordes said, but added that the company has no plans to commercialize its robots any time soon.

In addition to the robotic arm, Facebook has experimented with finding ways for a six-legged robot to teach itself how to walk. It said it eventually hoped to reduce the time needed to train such a skill to hours instead of days or weeks.

With computer scientists from UC Berkeley, Facebook has experimented with having robots learn about their world through tactile sensors that give them “a sense of touch” rather than relying solely on computer vision.

Yann LeCun, Facebook’s chief AI scientist, said in an interview Friday that the company’s researchers had a responsibility to “see around corners” to where technology might be heading. If robots do eventually become a popular consumer good and Facebook decides to sell them, he said, the company would need to have experts already on staff.

“You’ve got to start early,” he said. “It’s not just something you can jump into when it picks up.”