Even in spite of the fact that modern robots are very good at handling various kinds of objects, their program must first include data on these very objects. And if a robot comes across something with which it does not know how to handle it, it drives the car into a stupor. However, a group of researchers from the Massachusetts Institute of Technology (MIT) has created an algorithm that will train robots to interact with objects that they have never encountered before.
According to Engadget, engineers from the MIT Computer Science Laboratory are responsible for the development. And their invention could make robots more useful not only in production but also at home. At the moment, the main type of software embedded in robots allows you to execute an extremely limited set of commands: for example, select among objects (previously known to the program) the one that is needed at the moment or move objects pre-written in the algorithm from place to place.
A new AI from MIT, dubbed DON (Dense Object Nets), will allow robots to determine the structure and purpose of the object they are seeing at the moment. For example, a sneaker needs to be placed on the sole, and it is more convenient to hold the mug by the handle and preferably not to turn it over, as it may be filled with liquid. Based on the data included in the program, the neural network can independently improve, learning new subjects for itself. Moreover, AI remembers not only specific objects, but also their purpose. For example, if he is given the task of studying a subject, the AI will take several pictures and compare them with the available ones. Then the object is virtually decomposed into "component parts", the frame of the object and its most important points are drawn up. After that, it will be possible to interact with the object. Moreover,if you repeatedly give the task to find an object among the same similar ones, the neural network will not only immediately identify it, but also remember the whole class of similar objects, realizing that you can interact with them in the same way as with the original one.
Vladimir Kuznetsov
