Researchers at the University of California have developed an algorithm that teaches robots to walk without requiring human participation in this process.
In the course of the experimental part of the study, the autonomous system successfully trained the four-legged Minitaur robot to cross both familiar and unfamiliar terrains, while this method of learning either did not yield to traditional solutions in terms of movement efficiency or exceeded them.
Our development opens the way to creating patterns of movement, ideally adapted to a specific robot of any configuration and to a specific landscape; providing the best maneuverability, energy efficiency and reliability ", - engineers say.
Reportedly, the algorithm works on the principle of learning with reinforcement. He "punishes" the robot for large angular accelerations and angles of inclination and "rewards" - for moving forward. It took Minitaur only two hours to learn how to walk fully. In the process of learning, he made 160,000 steps.
According to the engineers, the technique turned out to be so effective that ultimately the device managed not only to overcome the flat surface and step over the cubes it had learned but also to climb the slope and steps - although the robot saw them for the first time.
The researchers emphasize that the presented solution is characterized by complete autonomy: the robot does not need preliminary computer simulation, it learns independently and only in practice.
The only drawback of this approach is the fact that the robot in the learning process will repeatedly make mistakes and lose balance, which is why its body must be strong enough, and its equipment must be shock-resistant.