Researchers develop system enabling self-supervised robot learning

A team of researchers from the Berkley University and the Technical University of Munich have developed a system that enables an artificially intelligent robot to perform self-supervised learning. In practical terms, the researchers built a robot that can predict that it will see through a camera if it performs certain movements. Instead of acting on the basis of pre-programmed instructions, the robot learns by itself, using unguided exploratory sessions to understand how things around it work. In the experiment, the robot was allowed to play with objects on a table; through deep learning, the robot was able to foresee how an image’s pixels would more from one frame to another based on its movements. This self-acquired understanding allowed the robot to move objects it had not dealt with before. Otherwise put, the robot was able to predict how different behaviours will affect the objects around it. At the moment, the system is basic and can only ‘predict’ a few seconds in the future. But a more advanced self-taught system could, for example, learn the layout of a factory, and have the foresight to avoid human workers and other robots who may be in the same environment, as Gizmodo explains.