Facebook announces ‘Habitat’, a platform for embodied ArtificiaI Intelligence research

Today, the Facebook research team announced ‘Habitat’, a new platform for embodied AI research. According to the team, this is a “modular high-level library to train embodied AI agents across a variety of tasks, environments, and simulators”. This will empower a shift from ‘internet AI’ with static datasets to an embodied AI model with agents acting in realistic environments.

The project was launched by the Facebook Reality Lab, Georgia Tech, SFU, Intel, and Berkeley to restore the disconnect between ‘internet AI’ and ‘embodied AI’. It will standardize the entire ‘software stack’ for training embodied agents, and release modular high-level libraries to train and deploy embodied agents. An important objective of Habitat-API is to make it easy for users to use a 3D environment and set up a variety of embodied agent tasks in it.

Habitat consists of Habitat-Sim, Habitat-API, and Habitat Challenge.

#1 Habitat-Sim

This is “a flexible, high-performance 3D simulator with configurable agents, multiple sensors, and generic 3D dataset handling”. It also has built-in support for SUNCG, MatterPort3D, Gibson and other datasets. Habitat-Sim achieves several thousand frames per second (FPS) running single-threaded, and reaches over 10,000 FPS multi-process on a single GPU on rendering a scene from the Matterport3D dataset.

#2 Habitat-API

Habitat-API defines embodied AI tasks, configuring embodied agents, training these agents, and benchmarking their performance on the defined tasks using standard metrics.