The trick was to implement a new training method that scaled well and stayed in sync no matter what the workload. Previous projects tend to struggle without massive computational power. Facebook taught a virtual agent to handle point-to-point navigation for the equivalent of 80 years of human experience — that’s about 2.5 billion steps. The result is an algorithm that, in indoor environments, is smart enough to choose the right fork in the path and quickly recognize errors when it does head in the wrong direction. It’s learning to understand the “structural regularities” of buildings, Facebook speculated.
The technology is still very young. It has yet to handle outdoors or complex situations, and it doesn’t handle long-distance navigation well if it has to lose sensors. Facebook is sharing its work in hopes of further advances, though. If that happens, it could not only help robots move gracefully from room to room, but help with augmented reality glasses and other systems that help you navigate unfamiliar spaces.