Why this matters Physical-world robotics needs high-quality, reusable digital homes to scale training and evaluation. This dataset supplies simulation-ready household environments with per-object physics, visual materials, and standardized trajectory recordings so teams can run large-scale embodied experiments without repeated manual asset preparation.
What Sets It Apart
- Scene and asset fidelity: 500 CAD-based household scenes, 31,711 placed object entries, 4,462 rooms and ~36,602 m² of floor area; assets include geometry, transforms, bounding boxes, textures, roughness and metallic parameters.
- Physics-aware assets: per-object physical parameters such as mass, density, static/dynamic friction and restitution are provided and manually reviewed for household scenes—so interactions and contact behaviors are more realistic in simulators.
- Task-ready recordings: 1,000 robot trajectory episodes, 1,316,359 recorded frames (12.18 hours at 30 Hz), RGB-D at 640×480; trajectories and observations stored in HDF5 with JSON/JSONL metadata following an RLDS-like format.
- Practical formats and tooling: scenes packaged as USD/USDz with
<scene>_metadata.json, object/material records, and meta directories; task chunks organized with episode-level HDF5 files and meta manifests for easy ingestion by simulators and training pipelines. - Open but constrained license: published under CC BY-NC-SA 4.0—suitable for research and non-commercial development but restricts commercial reuse without separate permission.
Who It's For and Trade-offs
Great fit if you build or evaluate embodied AI systems (navigation, manipulation, long-horizon household tasks) and need large, physics-configured synthetic homes that work across robot embodiments (URDF-friendly). Use it to prototype sim-to-real pipelines, benchmark policies, or scale data-driven policy learning without asset engineering.
Look elsewhere if you require permissive commercial licensing, ultra-high-resolution RGB renders beyond 640×480, or live sensor noise models and photo-realism tuned to a specific real-world camera; the dataset focuses on sim-ready geometry, physics fidelity and task-scale recordings rather than photorealistic rendering pipelines.
Data & Format Notes
Key files include USD/USDz scene packages, per-scene metadata JSON, and per-episode HDF5 trajectories with companion JSONL manifests. The dataset follows common embodied data conventions to ease integration into simulators and RL training stacks.
