Developed by Google, a project used real dogs to teach a robot dog to walk. To do this, the researchers used motion capture technology to record animals walking, running and jumping.
With this information, engineers were able to create a simulator for a digital version of a dog. Then, the developed algorithm could be applied to the “robot dog” Laikago, created by the Chinese company Unitree Robotics.
By using motors and not muscles, the robot acts differently than a dog. Still, the system created by Google professionals is considerably advanced. With it, it is possible to reproduce behaviors very similar to real animals.
According to the engineers, with motion capture, Laikago learned to walk faster than the initial factory version. However, it still moves unsteadily and needs some adjustments.
Regardless, they believe the system understood the information obtained from videos of real animals. In other words, it was not the long lines of the original programming that “taught” the robot to walk like a dog.
An (almost) real dog
An article on the Wired website revealed that the method used by Google to improve artificial intelligence is similar to real animal training. Thus, random movements are simulated and the digital dog receives rewards if he performs them correctly, like a real dog. If he does it incorrectly, he receives an alert not to repeat the action.
Surprisingly, the algorithm can recreate very natural behaviors, just like dogs. Although Laikago does not have an anatomy identical to the real animal, researchers have already observed it chasing its own tail like a dog.
The next challenge for Google’s engineers is to apply a more complex system to a physical experiment. In the real world, movements simulated with a robot are not always as accurate as they are applied in the digital world, as there are many more variables involved.