Google has recently been working in the field of so-called "cloud robotics". This is a phenomenon when robots, having learned to independently perform any action, can share their "experience" with other robots, simply by transmitting information by any available communication method. This principle of teaching allows you to avoid the moment of reprogramming, or, so to speak, "retraining", when setting new tasks to the technique.
The essence of "cloud robotics" is as follows: it is based on neural networks that determine and store the sequence of actions performed, are responsible for the processes of automatism and information transfer. In general, for everything that we call experience. Robots based on neural networks can set any task, and the artificial brain will find its own solutions. In the future, when performing these actions several times, the robot will develop an optimal algorithm that it will be able to transfer to other machines, and they will use and improve it, not starting from scratch every time.
Scientists from Google Research have tested their algorithm on three types of robots that perform different tasks: opening doors, studying objects on a tray, and a modified version of the first experiment, when the robot was not trained independently, but was controlled by a person with the subsequent task of improving the skills.
In the first case, it took the car a lot of time to understand that to open the door, you need to grab the handle, turn it and push the door. But all subsequent robots used this algorithm, skipping the training moment.
In the experiment with the tray, the machines were left to themselves and for several hours they studied the causal relationships between objects (for example: a kettle - a cup - sugar: what to do with this is obvious only to us, the robots had to "learn").
Experiment number three, after training the robot by the operator, was at the mercy of the "collective consciousness", which quickly jointly found the optimal solutions, differing in different initial positions of the manipulators and the final result, which accelerated the manipulation.
The most interesting moment was when one of the robots was forced to open a door on which a completely different type of handle was installed. The machine did an excellent job.
Why is all this necessary, in addition to constructing theories about the uprising of machines? It's simple: this acceleration of the learning process will enable industrial robots to start performing complex tasks much faster than with the traditional approach.
Promotional video:
VLADIMIR KUZNETSOV