# Habitat — High-Performance 3D Simulator for Embodied AI

> A flexible 3D simulation platform from Meta for training embodied AI agents in photorealistic indoor environments with fast rendering and physics.

## Install

Save in your project root:

# Habitat — High-Performance 3D Simulator for Embodied AI

## Quick Use
```bash
pip install habitat-sim
pip install habitat-lab
```
```python
import habitat_sim

cfg = habitat_sim.SimulatorConfiguration()
cfg.scene_id = "data/scene_datasets/habitat-test-scenes/apartment_1.glb"
agent_cfg = habitat_sim.agent.AgentConfiguration()
sim = habitat_sim.Simulator(habitat_sim.Configuration(cfg, [agent_cfg]))
obs = sim.get_sensor_observations()
# obs["color_sensor"] contains the RGB image
```

## Introduction
Habitat is Meta AI's open-source platform for embodied AI research. It provides a high-performance 3D simulator (Habitat-Sim) and a modular training library (Habitat-Lab) for developing agents that navigate, interact with objects, and follow instructions in realistic indoor environments.

## What Habitat Does
- Renders photorealistic indoor scenes at thousands of frames per second on a single GPU
- Simulates agent navigation, object manipulation, and physics interactions
- Provides benchmark tasks including PointNav, ObjectNav, and Rearrangement
- Supports training with reinforcement learning, imitation learning, and zero-shot policies
- Loads 3D scene datasets including Matterport3D, Gibson, HM3D, and HSSD

## Architecture Overview
Habitat-Sim is a C++ core with Python bindings that renders scenes using OpenGL or Vulkan. It loads meshes from standard 3D formats (GLB, GLTF) and uses Bullet for rigid-body physics. Habitat-Lab layers on top as a PyTorch-based training framework with modular task definitions, observation spaces, and policy architectures. Batched simulation across multiple environments runs in parallel on a single GPU for efficient reinforcement learning data collection.

## Self-Hosting & Configuration
- Install habitat-sim via pip or conda with GPU rendering support
- Download scene datasets from the Habitat data repository
- Configure sensors (RGB, depth, semantic) and agent embodiment in YAML configs
- Set up multi-environment batched simulation for RL training throughput
- Use Habitat-Lab's task and policy configs to define custom training experiments

## Key Features
- Renders at 10,000+ FPS on a single GPU, enabling fast RL training loops
- Supports photorealistic scenes from real-world 3D scans (HM3D, Matterport3D)
- Built-in physics engine handles object interaction and rearrangement tasks
- Modular task system makes it easy to define new embodied AI benchmarks
- Powers the Habitat Challenge, an annual competition for embodied AI agents

## Comparison with Similar Tools
- **AI2-THOR** — Unity-based household simulator; Habitat focuses on photorealistic scanned environments with higher rendering speed
- **iGibson** — physics-heavy interactive simulator; Habitat prioritizes rendering throughput for large-scale RL training
- **Isaac Sim** — NVIDIA robotics simulator; Habitat focuses on embodied AI research rather than industrial robotics
- **Sapien** — articulated object manipulation simulator; Habitat covers broader navigation and rearrangement tasks
- **ThreeDWorld** — multi-modal simulation platform; Habitat offers faster rendering and larger scene datasets

## FAQ
**Q: What GPU is required?**
A: Any modern NVIDIA GPU with OpenGL 4.1+ support. A GPU with at least 8 GB VRAM is recommended for photorealistic scenes.

**Q: Can I use my own 3D scenes?**
A: Yes. Habitat-Sim loads standard GLB/GLTF meshes. Use Blender or other 3D tools to create or export scenes.

**Q: Is it only for navigation tasks?**
A: No. Habitat supports object manipulation, rearrangement, instruction following, and social navigation tasks.

**Q: How does batched simulation work?**
A: Habitat-Sim can run multiple environment instances in parallel on a single GPU, providing high-throughput data collection for RL training.

## Sources
- https://github.com/facebookresearch/habitat-sim
- https://aihabitat.org/

---
Source: https://tokrepo.com/en/workflows/asset-2af51985
Author: AI Open Source