The purpose of this study, which the company conducted with the University of California at Berkeley, was to find a way to effectively transfer dog movements to robots. This could have been done before, but traditional training methods require a lot of involvement of scientists who, after independent training of robots, must adjust each movement in order to teach devices a new skill.
The Google project partially solves this problem – as the company noted, they added “a little controlled chaos” to the training. To do this, they recorded every movement of the dogs, tracking key points – paws and joints. Then these movements adapted to the movements of robots in a digital simulation, which should simulate devices.
In addition, the researchers introduced an element of randomness into the physical parameters used in the simulation, making the virtual robot weigh more, or have weak legs, or experience more friction with the ground. This made the machine learning model take into account all kinds of small deviations and ways to deal with them.
Having learned to adapt to change, the new training method allowed robots not only to run more stably but also to perform complex movements – rotations and turns. At the same time, scientists practically did not interfere with the training of devices.