Habitat

Meta's open-source embodied-AI simulation platform, used mainly for indoor navigation, object rearrangement, and human-robot collaboration tasks.

Habitat is an open-source embodied-AI simulation platform developed by Meta (formerly Facebook AI Research) together with several universities, first described in a paper at ICCV 2019. It has two layers: Habitat-Sim is a high-speed 3D simulator that can load real indoor scans from datasets such as Matterport3D, HM3D, Gibson, and Replica, and can render over 10,000 frames per second using multiple processes on a single GPU; Habitat-Lab is the upper-level library used to define tasks, configure agents, and handle training and evaluation. Habitat 2.0 (2021) added physical interaction and household rearrangement tasks, and Habitat 3.0 (2023) added humanoid avatars and human-in-the-loop control, supporting social navigation and human-robot collaborative rearrangement. It is one of the main platforms for visual navigation research, and hosts an annual Habitat Challenge that requires participants to submit code for evaluation in scenes it has never seen.

ExampleIn a point-goal navigation task, an agent is placed in a previously unseen apartment scan from HM3D and told only the target's coordinates relative to its starting point; it must reach the target using nothing but its camera feed and depth map.

Also called

Habitat-Sim, Habitat-Lab, Habitat 3.0, AI Habitat

Related

Point-Goal Navigation · Object-Goal Navigation · Habitat-Matterport 3D Dataset · Social Navigation · Rearrangement · Matterport3D

Sources

AI Habitat 官网 (Chinese)
Habitat: A Platform for Embodied AI Research (arXiv 1904.01201)
Habitat 3.0: A Co-Habitat for Humans, Avatars and Robots (arXiv 2310.13724)

As of

2023-10