Principle:Facebookresearch Habitat lab Environment Setup
| Knowledge Sources | |
|---|---|
| Domains | Embodied_AI, Simulation |
| Last Updated | 2026-02-15 02:00 GMT |
Overview
A prerequisite step that provisions 3D scene datasets and navigation meshes required for embodied AI training and evaluation in simulated environments.
Description
Environment Setup involves downloading and organizing the 3D scene assets (meshes, textures, navigation meshes, and episode definitions) that the Habitat simulator requires to instantiate environments. Without these assets, no training or evaluation can proceed. Habitat supports multiple scene datasets including Gibson, Matterport3D (MP3D), HM3D, and ReplicaCAD, each with different licensing and quality characteristics.
The setup process downloads compressed archives from Meta's dataset servers, extracts them into a standard directory structure under data/, and validates that the required files (`.glb` scene meshes, `.navmesh` navigation graphs, and JSON episode files) are present.
Usage
Perform this step before any training, evaluation, or interactive session that requires simulated 3D environments. This is always the first step in any Habitat workflow.
Theoretical Basis
The concept follows the data provisioning pattern common in ML pipelines:
- Asset identification: Select dataset UIDs matching the target task (e.g., `habitat_test_scenes` for testing, `gibson` for PointNav, `replica_cad_dataset` for rearrangement)
- Download and extraction: Retrieve compressed scene archives from remote servers
- Directory structure validation: Ensure files are placed in the expected `data/` subdirectory hierarchy
- Episode dataset linking: Verify that episode JSON files reference valid scene files