In order to robotized mechanism learned to walk, you can’t just “attach” to it a few feet. Training the movement is a very complex process that requires the developers a lot of time. But now this question will solve the artificial intelligence, because the group of experts had created a universal algorithms that help the AI learning robots any configuration to move. The human intervention in this process is not required.
The development is a team of scientists from the University of California at Berkeley and the group of experts Google Brain, one of the research divisions of Google on artificial intelligence. Their new system has trained four-legged robot to cross the terrain as familiar and unfamiliar.
“Deep reinforcement learning can be used to automate a number of tasks. If we can teach the robot gait from scratch in the real world, we can create controllers, which are ideally adapted to each robot, and even to individual landscapes, allowing you to achieve the best maneuverability, efficiency and reliability.” — said the scientists.
Reinforcement learning is essentially a method of carrots and sticks tailored for AI. It uses a reward or punishment in achieving or not achieving goals.
“Deep reinforcement learning is widely used for training the AI, and even for data transfer to the real robots, but this inevitably entails some performance loss because of inconsistencies in the modelling and requires active intervention. The use of such algorithms in real time was a difficult task.”
For experiments, the scientists took the robot Minitaur. They have developed a system consisting of a workstation, which is updated data of the neural network, downloaded information in Minitaur and unloaded back. Chip NVIDIA Jetson TX2 on Board of the robot was responsible for processing information. The robot walked for 2 hours and made 160 000 steps. During this time, the algorithm is rewarding the robot for moving forward and was punished if he was stuck in place or have made a very big roll to the side. In the end, was created a movement algorithm that allowed the robot to choose the optimal trajectory.
“As far as we know, this experiment is the first example of the application of reinforcement learning, which allows you to teach the robot to walk.”
Even more interesting and exclusive news on the topic of AI and high technology you can read in our news channel in Yandex.Zen.