A group of researchers at Google AI lab have developed a system that allows a robot dog to learn the most accurate copying of the movements and behavior of a real dog by “viewing” specially prepared videos. Note that teaching robots to perform certain tasks, imitating the movements of living beings, has long been nothing new. For example, industrial robotic manipulators that assemble cars on a conveyor belt, perform welding work and tighten bolts, copying the movements of human hands.
However, the industrial robots mentioned just above operate either according to a pre-created program or are trained using special devices. They cannot be trained by analyzing a video on which a person’s actions are captured, and this function is an innovation developed by Google AI lab specialists. In this case, the four-legged Laikago robot learns to walk, move its body and even use its tail with the help of video recordings of the movements of a real dog.
In practice, the training video is first processed by an artificial intelligence system that recognizes movements, adapts and checks them using the Laikago virtual robot. To eliminate possible errors of interpretation and adaptation, taking into account the limited capabilities of mechanical joints and drives of a robot dog, the artificial intelligence system must analyze several videos, preferably taken from different angles. As a result of this, the system creates a database of movements, linking them with certain behavior scenarios. And as soon as this base is tested on a virtual robot, it is loaded into the “brain” of a real robot.
The video below shows in detail the operation of this technology. Of course, it is very easy to notice some of the shortcomings of such a training system, especially compared to more advanced robots that get their skills through programming, for example, robots of the famous company Boston Dynamics. And this difference is especially noticeable in “unpleasant scenarios”, when the robot needs to restore the gait after “tripping” over an obstacle or to rise after a fall.
But researchers at the Google AI lab have a fair amount of optimism about this. They expect that the increase in the functionality of their system, which they planned in the near future, will allow the robot not only to get more natural behavior, but also to gain the ability to independently search for ways out of emergency situations.