The artificial intelligence models inside machines such as industrial robots require the ability to interact safely and efficiently with their environment. But training an AI in a real-world setting, even a risk-free training space, is often impractical because of the colossal amount of time required to teach complex neural networks.

To help speed up the process, Facebook Inc. today open-sourced a platform called AI Habitat. It’s a simulator that enables researchers to carry out many training sessions at once in a virtual environment.

AI Habitat might not be the first simulator built with machine learning projects in mind, but it’s certainly one of the fastest. According to Facebook, internal tests showed that the platform can run virtual training environments at a rate of 10,000 frames per second using just a single graphics card. The other simulation engines that the social network tried achieved just 50 to 100 frames per second.

Manolis Savva, Abhishek Kadian, Oleksandr Maksymets and Dhruv Batra, four of the engineers who worked on Habitat AI, wrote in a blog post that their team “pushed the state of the art in training speeds, making the simulator able to render environments orders of magnitude faster than previous simulators.” Facebook claims that experiments which take months with other tools can be completed in just hours using AI Habitat.

The company built the platform to be usable in a broad range of AI projects. Facebook made it modular, so that researchers can swap out key components if they need to, and implemented native support for several common open-source datasets used to train AI models.

The social network has open-sourced a training dataset of its own that is specifically designed to be used together with AI Habitat. Dubbed Replica, the repository includes 18 virtual indoor environments ranging from an office conference room to a two-story home.

Replica is so named because it “sets a new standard in the realism and quality of 3-D reconstructions of real spaces,” explained Julian Straub, one of the Facebook researchers who worked on putting the virtual environments together. They’re not renderings created in a graphic design tool, but scans of real-world spaces that the social network took using custom camera technology.

The company harnessed an infrared depth capture system to record the exact shape of the objects in each environment. Items such as books, chairs and windows were then reconstructed in high resolution, complete with their textures, to make sure that AI models are exposed to realistic-looking training environments. Facebook’s researchers also added in labels describing each object.

Researchers can use AI Habitat to train neural networks in performing fairly complex, multistep tasks that require the ability to not only see but also understand their surroundings. A domestic robot, for instance, could be equipped with the ability to handle requests like “check if my laptop is on my desk in the kitchen.” But the training environments in AI Habitat are not interactive, meaning it’s impossible to teach the robot how to fetch the laptop once it has found it.

Facebook plans to change that in the future. “The platform’s features will allow us to incorporate physics-based interactions, so mobile agents can manipulate 3D objects,” Savva, Kadian, Maksymets and Batra wrote. “But the full impact of AI Habitat will depend less on our upgrades than on its adoption.”

Facebook hopes that the platform will help researchers make progress towards developing next-generation autonomous machines such as robotic home assistants that require the ability to visually navigate complex environments. The social network also sees AI Habitat lending itself to augmented reality applications. An AR rendering algorithm trained in true-to-life simulated spaces, like those provided by the Replica dataset, might be capable of generating more realistic graphics.

Image: Facebook