Facebook has created a number of open source AI models that, according to the company, have an accuracy of 99.99 percent. The models are designed to allow robots to navigate indoor spaces without the use of maps.
The models should eventually lead to a way to create smart robots and voice assistants that can also take the spatial layout of a particular room into account. AI systems such as industrial robots need a map of the environment in which they operate to prevent them from taking a wrong turn, for example. The problem is that in many cases it is impossible to obtain a perfectly updated layout of the room.
“Most real-world environments evolve – buildings change, objects are moved, and people and pets are constantly on the move,” Facebook researchers Abhishek Kadian and Erik Wijmans explain.
The goal is to equip robots with the ability to independently navigate through an environment using only their sensors. Facebook claims that the models it open-sourced this morning are a big step toward that goal. In tests, the algorithms managed to achieve the goal 99.9% of the time using only a camera, a compass and GPS.
Kadian and Wijmans argue that the performance basically solves the Habitat Challenge, a competition launched by Facebook last year to accelerate research into autonomous navigation. The previous top performance AI achieved an accuracy of 92%.
Facebook has developed the AI models with a new training method called DD-PPO which is also shared with the open source community. The next goal is to develop an even more advanced model that can achieve comparable results using only a camera, without the need for compass or GPS data.