Habitat-Matterport 3D (HM3D)

Habitat-Matterport 3D (HM3D) is the largest and highest-quality dataset of real-world 3D indoor spaces for embodied AI research, jointly created by Meta AI, Matterport, and academic partners. The dataset provides 1,000 high-resolution 3D environment reconstructions from real Matterport scans, covering 215 million square meters of navigable space across diverse indoor environments. Each environment includes semantically annotated objects, regions, and surfaces, with high-fidelity geometry and textures. HM3D serves as the backbone for the Habitat platform's core tasks including point-goal navigation, object-goal navigation, exploration, and rearrangement. HM3D significantly improved over prior datasets (Gibson, MatterPort3D) in terms of scale, quality, and diversity. It has become the de facto standard for evaluating embodied agents in photorealistic simulation, and is widely used in conjunction with the Habitat simulator for training navigation and interaction policies.