Habitat-Lab — Embodied AI Training Framework

Habitat-Lab — Embodied AI Training Framework

Habitat-Lab is Meta FAIR's high-level library for embodied AI research, providing modular components for training agents on navigation (PointNav, ObjectNav), manipulation (rearrangement, pick-and-place), and social tasks inside the Habitat-Sim photorealistic simulator.

Key features

• Task suite: PointNav, ObjectNav, Social Nav, Pick, Place, and Rearrangement
• Modular architecture: sensor suites, reward shaping, agent configs via YAML
• VectorEnv for massively parallel rollout collection (100+ envs on one GPU node)
• Integration with DD-PPO (distributed RL), habitat-matterport3D datasets
• Evaluation on the standard SPS (steps-per-second) and SPL (success weighted by path length) metrics

Quick start

pip install habitat-lab

import habitat

env = habitat.Env(config=habitat.get_config("pointnav/pointnav_habitat_test.yaml"))
obs = env.reset()
while not env.episode_over:
    action = env.action_space.sample()
    obs = env.step(action)
metrics = env.get_metrics()
print(f"SPL: {metrics['spl']:.3f}")

npx ai-supply add habitat-lab-embodied-ai

Curated mirror of the open-source Habitat-Lab (MIT). Get it from the source.