Google Researchers Use AI to Teach Robots to Move like Real Dogs

They achieved this impressive feat by using something called reinforcement learning (RL). They began by taking a reference motion clip recorded from an animal and using RL to get the robot to imitate those motions.

"By providing the system with different reference motions, we are able to train a quadruped robot to perform a diverse set of agile behaviors, ranging from fast walking gaits to dynamic hops and turns. The policies are trained primarily in simulation, and then transferred to the real world using a latent space adaptation technique that can efficiently adapt a policy using only a few minutes of data from the real robot," wrote the researchers in their blog.

Randomness

However, it is a well-known fact that simulators provide a poor approximation of the real world, meaning that simulations don't perform well in reality. This is where the researchers decided to use a sample-efficient latent space adaptation technique.

They did so by introducing an element of randomness to the physical parameters used in the simulation by varying physical quantities, such as the robot’s mass and friction. This resulted in a machine learning model that could account for all kinds of small variances and the complications they create down the line.

The end result is a robot that moves with the same agility as a real dog. This kind of work is crucial as it can open opportunities to deploy robots for sophisticated tasks in the real world.