At the moment, no matter how capable a robotic mechanism is, in most cases it acts according to pre-written algorithms (or is completely controlled by a person) and cannot foresee the consequences of its actions. But scientists from the University of California at Berkeley have created a new technology for learning machines that allows them to predict where this or that action will lead, even when interacting with objects that the robot has not previously encountered.
The EurekAlert edition writes about the new development. According to their reports, the technology is arranged as follows: at the first stage, without the participation of an operator, using images received from cameras, the robot independently moves objects, on the basis of which a predictive model of behavior is built. After several attempts, it is "remembered" and used later when meeting new or similar seen objects. The robot's cameras calculate the possible sequence of actions, but so far, however, only a few seconds ahead. Although this is already enough for the robot to make decisions in new conditions for it without human participation. As one of the researchers, Frederick Ebert, “Throughout their lives, people learn to interact with the world around them practically unaided by trial and error, gaining the necessary experience and acquiring skills. We have managed to build a robotic system that will use a large amount of autonomously collected data to acquire useful skills, and the knowledge gained during training can be used by the machine in the future when, for example, moving objects."
Vladimir Kuznetsov
