RoboTHOR Sim-to-Real Embodied Navigation Platform
Provider：Allen Institute for AI (AI2)
RoboTHOR provides 89 apartment scenes and 731 object instances, some physically built in the real world, purpose-built for studying sim-to-real transfer.
At a glance
- Readiness
- L1 · Rigid-body ready
- Rationale
- Manipulation L1: 89 modular apartment scenes and 731 unique object instances, with part of the scenes physically built in the real world; assets carry collision and interactable semantics but no material-level physical annotation, hence L1.
- Scale
- 89 apartments (60 train, 15 val, 4 test-dev, 10 test-standard), 11 furniture types and 32 small-object types, 731 unique object instances (The RoboTHOR paper (arXiv 2004.06799) reports 89 apartments and 731 unique object instances; the 2021 challenge splits them 60/15/4/10. Checked 2026-10-25.)
- License
- Apache-2.0 — The AI2-THOR framework and images generated with it follow Apache 2.0; RoboTHOR scenes and assets ship with the same platform, so verify per-asset provenance before commercial use.
- Access
- Open
- Formats
- other
- Runtimes
- other
- Physics properties
- Collision geometry
- Domain
- Other
- Published
- 2026-10-25
- Updated
- 2026-10-25
Why it is listed
RoboTHOR exists to answer the question most simulated benchmarks avoid: does a policy trained in simulation survive contact with a physical apartment? Its scenes are hand-built by 3D artists and then partly reconstructed as real rooms near the Allen Institute in Seattle, so simulation and reality can be compared on the same layout.
Readiness assessment
Eighty-nine modular apartments are split into sixty for training, fifteen for validation, four test-dev and ten test-standard, built from eleven furniture types and thirty-two small-object types across 731 unique instances. Fourteen object categories are designated navigation targets and guaranteed present in every scene, and layouts are deliberately non-overlapping across splits so a policy cannot memorize geometry.
Practical notes
The framework runs through the AI2-THOR Unity simulator with agents including LoCoBot, StretchRE1, ManipulaTHOR and a drone; the physical robot is a LoCoBot with a RealSense camera, mirrored in simulation with matched camera properties and controller noise.
Disclaimer: license and usage terms follow the official source — verify them yourself before use. This site indexes and links only; it does not host or mirror any asset file.
License and usage terms are those of the official source — verify before use. No asset file is hosted or mirrored on this site.