Google robot teaches itself to walk

Hyperaxion Mar 4, 2020

A year after researchers discovered how to make robots learn in a real environment, in a few hours Google’s Rainbow Dash robot learns to walk.

Artificial intelligence is revolutionizing everything from education to medicine, architecture and even economics, and this process can be accelerated with artificial intelligences capable of teaching themselves.

A Google robot called Rainbow Dash did just that. Within hours, it was able to teach itself how to walk. Even without substantial help from humans, Rainbow Dash was able to learn to walk on more than one type of surface.

Deep reinforcement learning

According to an article published on the ArXiv preprint server, the researchers used a technique called deep reinforcement learning to achieve this.

With this technique, learning occurs through trial and error and is usually employed in games, in which the artificial intelligence learns how to win.

It is unusual for deep reinforcement learning to be used in the real world, outside a simulated environment. Since learning happens through trial and error, in a virtual environment there is no risk of accidents or damage from falling.

Read more: AI identifies a new antibiotic effective against superbugs

Deep reinforcement learning usually starts in a virtual environment and the code is subsequently transferred to a robot, thus preserving its safety. This technique differs from others in that it is not necessary a previous set of data to teach the AI.

Did the researchers intervene?

The researchers intervened by writing an algorithm to prevent Rainbow Dash from falling, as well as placing barriers to prevent the robot from leaving the learning area. The algorithm written by the researchers focused on containing the robot’s movement.

According to Chelsea Finn, a professor associated with Google, removing human beings from the learning process is complicated. He adds that by allowing robots to learn for themselves, they get closer to learning in the real world.

Indeed, to adapt to the real world, an artificial intelligence must learn to deal with a multitude of adversities. The ability to teach itself can make this process more efficient.


Leave a Reply

Notify of