Facebook researchers have developed a deep learning algorithm capable of navigating the real world with nearly 100% accuracy, reads Facebook press release from January 21.
The innovative algorithm called DD-PPO is a new large-scale distributed reinforcement learning (RL) algorithm, that uses RGB-D camera, GPS, and compass data to learn about surroundings and effectively navigate through it without a map.
Researchers have reported an unprecedented level of accuracy – agent DD-PPO reached the desired goal in 99,9 % of the time, during the tests inside the houses and office buildings.
Researchers also report that “the method scales well”, while the navigation paths chosen by AI remain intelligent. Additionally, in case of creating an error, for example induced by a mirror, it recognizes it and corrects the path.
Effective interaction has been a long-standing challenge for robotics. While navigation with maps has been achieved a while ago, real world offered many challenges, with maps being “outdated the moment they are created”. Therefore, this achievement marks an important step in further applications of technology:
“By learning to navigate without a map, DD-PPO-trained agents will accelerate the creation of new AI applications for the physical world”
The technology could be especially useful for robotic assistants, broadly widening their functionality. On the other hand, Facebook team takes a new challenge – to create an algorithm that will not depend on compass or GPS data, but will be able to analyse surrounding only by camera input.
“We are creating a new challenge to perform point-goal navigation using only RGB-D input”