It may be a robot, but that’s not why it ceases to have a name. This ‘four legged’ machine is called Jueying and is the cover highlight of the latest edition of the publication Science Robotics. What’s so special? He is learning, on his own, to respond to the ‘chaos’ of the world around him – whether it is walking on terrain full of stones or defending himself from kicks and shoves from humans.
The project is a joint investigation by universities in Zhejiang, China, and Edinburgh, Scotland. The great differentiating element of this robot dog is how the team is taking advantage of different reinforcement learning algorithms to learn how to adapt to the conditions that surround it.
Each of these algorithms learns a specific skill – such as trotting or getting up if you fall on your back – and is compensated or portrayed – digitally -, respectively, if you do something that is considered ideal or not ideal. These algorithms are trained first in a digital simulator and only then integrated into the hardware itself.
As the magazine explains Wired, Jueying has eight distinct algorithms integrated, each with the objective of helping the robot to produce complex behaviors that make it move and overcome obstacles. These algorithms are then linked together through a central system, which uses, in real time, a combination of the response of the different algorithms so that the dog can overcome these obstacles.
That is why when the robot dog is kicked or pushed, it manages to have a coordinated response between the different algorithms to create a defense response. If kicked, the robot will roll after falling, so it can get up immediately. If you are traveling on a path with stones and fall, you will get up and continue the march, adapting to the uneven terrain.
Researchers consider this to be the most suitable software approach for robots, as it is very difficult to code line by line all possible responses that the robot can give to a given obstacle. With this method, the algorithms learn to be more effective and work together to achieve the best possible result.
“The Artificial Intelligence approach is very different in that it captures an experiment, which is based on hundreds of thousands of attempts by the robot, or even millions of attempts,” explained robotics researcher Zhibin Li. simulated environment, I can create all possible scenarios. I can create different environments or different configurations. For example, the robot can start in a different position, like lying on the floor, standing, falling and so on ”.
According to the expert, “we will have smarter machines, which are able to combine flexible and adaptable skills in real time, to respond to a variety of tasks that they have never seen”.