Facebook robots no longer need maps for navigation

Facebook has developed a new learning algorithm that allows robots to navigate in space without the use of maps.

Although robots could mix beforeon their own, but at the same time they had obvious problems with navigation, which led to returns or even falls when they collided with an unexpected obstacle. However, thanks to the new Facebook solution, robots can without a map reach their destination in 99.9% of cases and with only a 3 percent deviation from the ideal route.

At the same time, in order to effectivelythe developed DD-PPO (decentralized distributed optimization of proximal policies) algorithm works with a fairly standard RGB-D camera, GPS and compass.

The team achieved success through trainingthe car to lay a route between two points, using data on movements from 2.5 billion steps or about 80 years of human experience. This makes the program smart enough to choose the right turns on the way and quickly identify errors when driving in the wrong direction.

Despite success in understanding structurallaws of buildings, technology is still in its infancy. So far, it is difficult for her to cope with navigation on the street and in difficult situations, as well as paving the way for long distances in case of loss of communication with sensors.

Facebook shared the results of the study,since it expects to significantly improve the capabilities of the algorithm. If successful, the technology will not only increase the efficiency of robotics, but also various smart devices that can help users navigate in unfamiliar places.

 In addition to improving navigation systems, Facebook also plans to teach machines to sense objects and manipulate them using tactile sensors.

</p>