Google spent $ 1.5 million to study methods for learning robots like babies

Until now, the robot's actions and control abilities were only on par with that of a two-year-old.

Although robots are very good at watching, listening and learning about the world around them, they are still quite useless when interacting with things in a way that people do. They "struggle" when opening doors, walking downstairs and eating with knife and fork. "Their actions and controls are only on par with a two-year-old," said Abhinav Gupta, assistant professor of robotics at Carnegie Mellon University.

The robot is learning to grasp the object.(Photo taken from the clip).

Gupta and his team of researchers are hoping to change that. Their method is to allow robots to play with objects , allowing them to explore and learn to hold themselves like a child.

"Psychological research indicates that if people are not exposed to what they see, their understanding of it will be limited," said Lerrel Pinto, a student who is preparing to get a Ph.D. Learning robots in Gupta's group said . "The interaction with the real world will make the robot develop" sharpness "in its intellect."

Learning robots grasp objects.

The Gupta team demonstrated their results at the European Conference at the Computer Vision event last fall. And that demo has helped them win a $ 1.5 million "scientific research award" in three years from Google. This money will be used to increase the number of robots that they put into testing more, thereby creating a richer database for this team to work.

"If data is collected more quickly, we can do more - create more different frameworks," compose "more algorithms" , Pinto said. A robot's "learning" can be shared to other robots.

They are learning an innovative way to speed up the "learning" of robots faster. One approach to this goal is to use this skill - push an object down - let the robot develop other skills such as grasping, picking things up.

Methods for learning robots by Gupta and colleagues.

Another direction is adversarial learning. It is for a robot to learn how to grasp things, and another robot tries to shake or snatch the object. Think of this approach similar to how an athlete is trained with an enhanced mode, or a parent teaches his child how to catch the ball by throwing hard shots.

Until now, robots trained by adversarial learning have shown marked improvement in their skills, especially compared to robots who cannot apply it.