What it is
100K simulation-ready tabletop scenes built from real internet images, with physical object properties and expert manipulation trajectories for generalizable robot learning. 100K tabletop environments from single in-the-wild frames, with ~1M object instances across 35K+ categories, scene assets/XML, physical properties, and collision-free expert trajectories.
Key results
- 100K unique tabletop environments
- ~1M object instances across 35K+ categories
- Includes scene GLB/XML assets and CC BY 4.0 license
Useful for training and evaluating generalizable tabletop manipulation policies in simulation from real-world scene layouts.