photo-rai.ru

The Google company conducts works in the field of so-called & #171 recently; cloudy робототехники». This phenomenon when the robots who studied independently to perform any operation, can share & #171; опытом» with other robots, having just transferred information by any available method of communication. Such principle of training allows to avoid the reprogramming moment, or, so to speak, & #171; переобучен褻 at statement to the equipment of new tasks.

In the first case at the car a lot of time for understanding of that left, for opening of a door it is necessary to undertake the handle, to turn it and to press on a door. But all subsequent robots used this algorithm, passing the training moment.

In an experiment with a tray of the car were provided to themselves and within several hours studied relationships of cause and effect between objects (for an example: teapot & #8212; cup & #8212; sugar: what with it to do, it is obvious only to us, it was necessary to robots & #171; учитьс¤»).

Why all this is necessary, besides creation of theories about a revolt of cars? Everything is simple: similar acceleration of process of training will give the chance to industrial robots to start performance of difficult tasks much quicker, than at traditional approach.

The moment when one of robots forced to open a door on which the handle absolutely other type was installed became the most interesting. The car perfectly coped with a task.

Experiment number three after training of the robot by the operator was assigned & #171; collective сознанию» which quickly in common found the optimal variants of the decision differing in various starting positions of manipulators and the final result accelerating performing manipulation.

Essence & #171; cloudy робототехники» is as follows: neuronets which define are its cornerstone and keep the sequence of the performed operations, are responsible for processes of automatism and information transfer. Generally, for all that we call experience. Robots on the basis of neural networks can set any task, and an artificial brain itself will find solutions. Further when performing these actions several times, the robot will develop an optimum algorithm which will manage to transfer to other cars, and those will use and improve it, without starting every time from scratch.

Scientists from the Google Research company tested the algorithm on three types of the robots who are carrying out different tasks: opening of doors, studying of objects on a tray and the modified option of the first experience when the robot studied not independently, and was operated by the person with the subsequent task to improve the received skills.