Researchers at the University of California, Berkeley, have created a file robot dog Able to learn, in 20 minutes, to walk even on terrain that is challenging for robots, such as a meadow, forest trail, or memory foam mattress. a Robot use algorithm It is called Q-Learning (generally used in simulation), which does not require training in a functional terrain model.
This type of machine learning used in Search It was called deep reinforcement learning, in which a robot receives rewards for every action it takes, depending on how well it meets pre-set goals.
The robot constantly repeats the experiments, comparing them with the previously obtained positive results, until it learns to walk on the ground.
This process is different from that used with most autonomous robots of this type, which learn to walk after being tested simulation of terrain or movements that have been programmed by humans. These bots often struggle when confronted environments Not knowing or unexpected obstacles.
Robot dog is good but still needs improvement
The robot dog has successfully learned to walk on its own, however, according to the researchers, they will need to improve the robot’s reward system to learn to perform other tasks.
“In some ways, it’s very similar to the way people learn,” team member Ilya Kostrekov, of the University of California, Berkeley, told the website.new world“.” Interact with some environment, get some benefits and think mainly about your past experience and try to understand what could have been improved. “
“I find it very impressive,” says Chris Watkins of Royal Holloway, University of London. “I’m honestly a little surprised that you can use something as simple as Q-Learning to learn skills like walking on different surfaces with so little experience and so quickly in real time.”
