Autonomous artificial intelligence teaches robots to walk

Autonomous artificial intelligence teaches robots to walk

Autonomous artificial intelligence teaches robots to move without human help

The possibilities of autonomous artificial intelligence are enormous. Humanity always finds a new use for it. In particular, AI is able to replace a person when training other machines. This is another step to ensure that people abandon the routine work, and were able to fully devote themselves to creative activities.

What methods are used by autonomous artificial intelligence when learning

If a robot simply attaches two or four legs, there is no guarantee that it will be able to use them correctly. To teach them to walk is quite difficult, this task requires perseverance and a huge amount of time that could be spend on more valuable things. Human had to look for solutions that can save him from this problem. And the specialists did. A group of scientists developed a universal algorithm. With its help, machines will be able to learn the movements with an automatic “teacher”, which is artificial intelligence.
The process takes place autonomously, the participation of people is not required.
The authors of the idea were researchers from the University of California at Berkeley, as well as employees of Google Brain project. This division of American IT giant is engaged in the study and improvement of AI. The joint development of professionals trained four-legged robot to walk both on familiar and unfamiliar location.

Automated learning by consolidation can be used for solving a huge number of different tasks. If the technology shows itself with different machines, it can be a step towards creating new controllers generations. Control methods will be adapted not only to the individual droids, but for the landscape. Thus, it will be possible to achieve maximum maneuverability and reliability of equipment.
Training with consolidation is de facto similar to “carrot and stick”, only AI uses it. It is able to encourage the “ward” for the implementation of the requirements or vice versa, “punish” him. Such a solution is widespread and to train the artificial intelligence. Now it is possible to download and to help to learn important information to the robots.
However, the researchers drew attention to the problems. First of all, they noted some decrease in efficiency due to differences in algorithms.
It is not easy to apply such models in practice, and professionals have a lot of work to eliminate these differences.
For the experiment, experts used a droid called Minitaur. The project team created a special system, the so-called workstation.
Its tasks included updating information in the neural network, loading information into the robot’s memory, as well as their reverse upload. Data processing in Minitaur is done by Nvidia Jetson TX2 processor.
During the study, the machine made 160 thousand steps, which took about 2 hours. AI encouraged the droid if it moved, and punished when it stopped for a long time or leaned in any direction. In the end, an algorithm was developed that made it possible to move and select the appropriate trajectory regardless of the situation.
Scientists of the University of California commented on the results. According to them, this case is the first in which it was possible to teach the robot to walk with using the method of training with consolidation.