Researchers at the University of California, Berkeley, have developed a robot that has the ability to learn as a young child, allowing you to predict the results. Called Vestri, the robot is capable of learning by itself without human supervision.
Teaching a robot to play
When young children play with toys, they are doing more than just entertaining themselves. Indeed, with each turn or throw, children learn about how the world works. When manipulating objects, young children learn how they respond by themselves and then judgments can be made about how these objects will likely behave in the future if they are used in the same way.
This great learning strategy, sometimes called "motor babbling", was emulated by American scientists in the Vestri robot. The technology in question, called "visual foresight", allows the robot to imagine what its next action should be and what might be the most likely consequences, and then take action based on the best results.
"Children can learn about their world by playing with toys, moving them, grasping them, etc. Our goal with this research is to allow a robot to do the same thing: to learn how the world works through autonomous interaction," he said. assistant professor at the University of Berkeley Sergey Levine and lead author of the study presented at the Neural Information Processing Systems conference. . "The capabilities of this robot are still limited, but its abilities are learned fully automatically and allow you to predict complex physical interactions with objects you have never seen before, based on previously observed patterns of interaction.
Scientists expect that the future, that technology could allow the cars that drive themselves to predict the roads ahead, but for now, at least, this "robotic imagination" is quite simple and limited, Vestri can make predictions only in several seconds in the future but even that is enough to help him find the best way to move objects around a table without disturbing obstacles Vestri chose the right path about 90 percent of the time.
The crux of this skill set is that no human intervention or prior knowledge of physics is required. Everything that Vestri learned, he does from scratch from unattended and unsupervised exploration: "playing" with objects on a table.
After training, Vestri can build a predictive model of his environment. Next, he uses this model to manipulate new objects he has never encountered before. The predictions occur in the form of video scenes that had not really happened, but could happen if an object were pushed in a certain way.
"In the same way that we can imagine how our actions will move objects in our environment, this method can allow a robot to visualize how different behaviors will affect the world around him," Levine said. "This can allow intelligent planning of highly flexible skills in complex real-world situations."
Because video predictions of Vestri depend on observations made autonomously by the robot through camera images, the method is general and widely applicable. This contrasts with conventional computer vision techniques that require human supervision to label thousands or even millions of images.
Next, the Berkeley researchers want to expand the amount of objects Vestri can play with, but also to improve the movements of his ability. Of creation. By expanding their repertoire, researchers hope to make Vestri more versatile and adapt to all kinds of environments.
"This can allow intelligent planning of highly flexible skills in complex real-world situations," Levine concluded.
Scientific reference: NIPS Conference.
Enjoy this article? Join more than 40,000 subscribers to the ZME Science newsletter. Subscribe now!