Humans learn from experience to not “do dumb things,” and Facebook chief AI scientist Yann LeCun is trying to create a version of that for robots, saying that systems that learn “models of the world” are our best shot at advancing artificial intelligence. Unlike a rewards/demerits-based reinforcement learning, Facebook’s tack is to instill curiosity, by giving the robot freedom to try new things. With New York University, Facebook also dramatically reduced the number of tries to teach a robotic arm to grasp an object.
Wired reports that the technique is self-supervised learning, in which the robot “tries new things and updates a software model, which can help it predict the consequences of its actions.”
AI research scientist Franziska Meier said that, in the learning process, even when the robot fails, “it gave us more data, and the variety of data we get by exploring like this is bigger than if we weren’t exploring.” University of Oslo roboticist Tønnes Nygaard built a four-legged robot that used this technique to learn to walk on its own.
Facebook’s project marries AI and robots, which up until now have been more isolated pursuits, and “robots have remained fairly dumb.” Researchers, however, are now using robots “as platforms to refine [AI] software algorithms.” If Facebook teaches a robot to “solve a series of tasks,” that information “might inform the development of AI assistants that can better plan a sequence of actions for you, the user. “
“If you solve it in one context, you’ll solve it in the other context,” said LeCun. In this way, “AI is making robots smarter, but robots are also now helping advance AI.” “You can’t cheat with robots,” explained LeCun. “You can’t have thousands of people labeling images for you.”
LeCun added that Facebook is pursuing robotics because the company thinks it will have implications for technologies like telepresence. Wired points out that Facebook “is already a hardware company, after all, what with the Oculus VR system and Portal, its video conference device.”
Bloomberg reports that researchers at Facebook and 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.” “The great thing about robotics is that it takes place in real time, in the real world,” said Facebook co-managing director of the AI research lab Antoine Bordes, who added that the company “began working on robots in the past year because it forced researchers to think creatively about how to make machine-learning more efficient.”
The lab is “finding ways for a six-legged robot to teach itself how to walk,” and hopes the training time can be reduced from weeks to hours. Facebook has also teamed with UC Berkeley researchers to experiment with robots learning via tactile sensors rather than just computer vision. The company said it has no plans to commercialize robots in the foreseeable future.