A robot from Google Robotics has learned to walk on its own. Video

[zakiyatasnim]

Using algorithms from Google engineers, a four-legged robot learned to walk in just a few hours — 12 months faster than the average person.
Scientists from Google Robotics have successfully launched a robot that is able to learn independently, without relying on simulation tests. Human participation in this process is minimized. The research process is described in an introductory text that the team posted online.

— a type of machine learning based on trial and error while mexico number data providing rewards for certain actions — the scientists were able to avoid programming every command related to walking, which made it possible to avoid creating simulations.

Reinforcement learning is typically used in virtual environments. However, building a simulation that could replicate the robot's movement on different surfaces proved to be complex and time-consuming. So the researchers decided to train their system in the real world.

To do this, they allowed their four-legged robot to move across different surfaces, such as a flat surface, a soft mattress, and a rug with various gaps. Within a few hours on each platform, it learned to walk forward and backward and turn left and right. At the same time, manual control and restart were not required once.



The developers also managed to reduce the need for human intervention in situations where the robot fell or tried to leave the training area. They forced the robot to perform several maneuvers at once and limited its movement within a given boundary. When it reached the edge of its territory, it recognized the boundary and walked back, and if it fell, it could get up on its own.

That is, the only human intervention in the process was that they added an algorithm that tells the robot to get up after falling, but they hope to eventually automate this part of the learning process as well.

Once the robot learned to walk, the engineers turned on a game controller that allowed them to manually move the robot, using the techniques they had learned to test the learning results.