Humanoid robots have taken a step forward, and are now even capable of learning a series of movements, such as the process of walking, practically on their own.
And researchers at the University of Berkeley have created a humanoid robot capable of relying on artificial intelligence techniques to learn to walk on its own.
For this type of robots to be completely valid, they must adapt to a multitude of surfaces and terrain, which is why they require many hours of prior training.
And researchers have relied on classic reinforcement learning to teach the robot how to adapt to changing terrain and needs.
To teach you how to walk, they started with a simulation running billions of scenarios in a high-performance GPU-based physics simulation environment.
Then this simulator algorithm rewarded actions that imitated human walking and punished actions that did not do it well.
When this task was perfected, it was transferred to any humanoid robot that had not previously been tested in the field.
In this way, the researchers implemented a transformer that learns from observations and actions by predicting the consequences of each of the actions taken.
In this way the humanoid robot He was able to adapt his actions to any terrain, even if he had never been through it before.
What surprised the researchers, It seems that the humanoid robot even learned to move its arms when walkingthat is, imitating the classic walk of human beings to maintain balance.
This is really impressive considering that this humanoid robot lacks sensors that help it perceive its surroundings.
Although the walk of this robot is not as refined as that of human beings and even seems jerky, the fact that it lacks sensors requires the machine to have to adapt to the terrain as soon as it touches it, without previously preparing.
In this way, with this algorithm transferred to the robot, it is capable of moving throughout the university campus without apparent help and is a good start to facilitate this type of training for future robots.
