Habitat: 3D Indoor Environments for Embodied AI
Photorealistic 3D buildings for robot navigation, rearrangement, and mobile manipulation. From Meta AI Research.
Key Stats
| Metric | Value |
|---|---|
| HM3D | 1,000 photorealistic building scans |
| Matterport3D | 90 building scans |
| Gibson | 572 building scans |
| Benchmarks | PointNav, ObjectNav, Rearrangement, Mobile Manipulation |
| Simulator license | MIT (Habitat-Sim) |
| Data licenses | Varies per scan dataset (research-only for most) |
What is Habitat?
Habitat is Meta AI's embodied AI research platform, consisting of a high-performance 3D simulator (Habitat-Sim, running at 10,000+ FPS) and curated sets of photorealistic indoor 3D environments. The platform hosts several annual challenges on tasks like point-goal navigation (navigate to coordinates), object-goal navigation (find a chair), object rearrangement (put objects in their correct places), and mobile manipulation (pick up objects and move them).
The 3D environments come from three scan datasets: HM3D (1,000 buildings, the largest), Matterport3D (90 buildings, high quality), and Gibson (572 buildings). These provide the photorealistic indoor scenes that embodied agents navigate and manipulate within.
Access
License note: Habitat-Sim is MIT licensed, but the 3D scan datasets require separate agreements and are generally restricted to research use.
Related datasets
- RoboCasa -- kitchen manipulation simulation
- ManiSkill2 -- manipulation benchmark
- Real-World RL -- real-world reinforcement learning