Google Robots Share Experiences to Learn Faster

Researchers at Google have been working on robot learning through shared experiences between multiple robots. This means that one robot’s learned experience can be shared with other robots through a computer’s hive mind. One robot using multiple learning experiences speeds up the process of learning to do complex tasks. Researchers used cloud robotics and deep neural learning technologies together to see how robots utilized shared learning and how they responded to three learning experiences.

In the first case, the robots completed a simple task—opening a door on its own. They learned to do this over time by adapting to small variations in the environment. Then, the data from each separate robot was pooled together in the neural network and an analysis of these experiences was sent back to the robots. Each robot then learned from the raw trial and error experiences of the other robots.

In the second scenario, the robots interacted with a box of household items. Like the first case, individual experiences were shared with all the robots. These experiences were used to generate a predictive model, which allowed the robots to forecast what might happen given their actions. This lets them make purposeful movements in relation to the objects, such as moving a can of soda to a designated spot on a table.

In the last scenario, the robots learned to open a variety of different doors—with the help of humans. A human guided each robot to open a door. The information from the robot opening the door was encoded into a neural network where camera images are translated to robot actions, with a single strategy being applied for all robots. The robots then utilized the strategy and learned by trial and error for increasingly complicated variations, like doors being angled differently or having different handles. The more iterations, the more the strategy improved. Since they were trained to open the door by a human first, the robots were able to learn more effectively.

The robots showed that they can learn quickly through shared experiences and with human intervention. With more time and training, we may soon have robots that can help us out around the house or on factory floors.

Via: Google Research Blog

Follow us on Twitter for updates and news.

Related Articles