Facebook creates a “mirror world” for learning artificial intelligence

Dreams of a robot-Butler seem similar — I think here, within arm’s reach, and he rushes to open the door for you and wash your socks. But first we need to teach robots to navigate in the busy stuff of the human environment. For this Facebook has created an incredibly realistic simulator for the training of artificial intelligence every action in the room.

Virtual environments — from Atari games to simulator self-driving cars — have become a leading tool for teaching machines how to interact with the environment. One of the main reasons is that such environments give us full control over the environment of the agent. The real world is too downtrodden place with a huge number of variables that must be considered, but in the simulation we not only know all the variables — we can customize them at will.

It’s worth it. Controlled world simulation is very different from the environment in which it will subsequently operate a physical robot, so the agents have to practice thousands or millions of times to make it right. Virtual environments can be scrolled hundreds of times faster than real-time, and for parallel operation of hundreds of experiments requires only a few computer chips and not hundreds of new robots and test environments.

A new solution to this problem is to create more realistic digital environment. The point is that the closer we can simulate the real world in the simulation, the easier the AI will tolerate the fact that exploring in the real world. This is the logic inherent in a new set of database Replica from Facebook: photo-realistic 3D mirror for different internal areas including apartments and a retail store.

Replica from Facebook: how to teach artificial intelligence

Replica data was extracted directly from the real world using specially developed camera that combines HD video, high-precision data about the depth and simultaneous localization and mapping (SLAM) that records where all these measurements were taken.

This approach allows us to reflect even reflective surfaces and complex textures, which are often difficult to reproduce. In addition, Facebook engineers have carefully labeled everything in the data set, including color coding of different categories of objects. After that, all data can be downloaded to the platform-simulation of Habitat to create a three-dimensional environment that can be explored by AI.

The company not only helps the AI with navigation, it also hoped that the dataset will be able to improve the technology of augmented reality (AR). If AI will be able to better understand the physical space occupied by people, it would be better to combine digital projection with the real world.

Although not a perfect Replica, this is the perfect mirror of the human world that’s ever modeled. And Facebook will almost certainly have plans for expansion and enrichment of the dataset. I hope this will bring us to the dreams of robot butlers.

Or not? Let’s discuss in our chat in Telegram.

Leave a Reply

Your email address will not be published. Required fields are marked *