Habitat-Sim — High-Performance 3D Embodied AI Simulator

Habitat-Sim — High-Performance 3D Embodied AI Simulator

Habitat-Sim is Meta FAIR's high-performance 3D simulator for embodied AI research, underpinning Habitat-Lab. It renders photorealistic indoor scenes (Matterport3D, HM3D, Replica) with RGB, depth, and semantic sensors at 10,000+ frames-per-second on a GPU — orders of magnitude faster than Gazebo or Unity for RL training.

Key features

• 10,000+ SPS on a single GPU for massively parallel training
• Semantic instance / panoptic segmentation sensor
• Physics simulation via Bullet: rigid body, articulated robot control
• Loads URDF robots (Fetch, Spot, Stretch, LoCoBot, Franka)
• Python and C++ APIs; headless rendering on SLURM cluster nodes

Quick start

conda install habitat-sim -c conda-forge -c aihabitat

import habitat_sim

sim_cfg = habitat_sim.SimulatorConfiguration()
sim_cfg.scene_id = "path/to/scene.glb"
agent_cfg = habitat_sim.agent.AgentConfiguration()
sim = habitat_sim.Simulator(habitat_sim.Configuration(sim_cfg, [agent_cfg]))
obs = sim.reset()
print(obs.keys())  # ['color_sensor', 'depth_sensor', 'semantic_sensor']

npx ai-supply add habitat-sim-3d-simulator

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