Autonomous Learning

Google is creating AI-powered robots that navigate without human intervention—a prerequisite to being useful in the real world. It took 10 minutes for baby fawn to be able to stand and seven hours to walk. While autonomous robots are already familiar concept, autonomously learning robots are at the beginning of their development. Existing learning algorithms are still relying on human intervention.

The work is based on a project that took place one year ago, when the group discovered how to get the robot to learn in the real world. But a human still had to take care of the robot, and manually conflict  hundreds of times, says Jie Tan, a paper coauthor who leads the robotics locomotion team at Google Brain. “Initially I didn’t think about that,” he says.

So they began to solve this new problem. First, they restricted the area that the robot could explore and told him to train on many maneuvers simultaneously. If the robot reaches the edge of the boundary while learning to walk forward, it changes direction and begins learning to reverse.

Secondly, the researchers also limited the robot’s trial movements, making it careful enough to minimize damage caused by repeated falling. At a time when the robot inevitably collapsed, they added another encoded algorithm to help it stand up.

Thanks to these various modifications, the robot has learned to walk independently on several different surfaces, including a flat surface, a memory foam mattress and a wiper with slots. The work shows the potential of future applications that may require robots to move in difficult and unknown terrain without human presence.

The following two tabs change content below.

Grams Bartosz


Latest posts by Grams Bartosz (see all)

Leave a Reply