TableVerse

# TableVerse-100K: A Large-Scale Tabletop Dataset with Real-World Grounded Layouts for Generalizable Manipulation [](https://creativecommons.org/licenses/by/4.0/) [](https://github.com/google-deepmind/mujoco) [](https://huggingface.co/) [](https://huggingface.co/papers/2607.21017) [](https://arxiv.org/abs/2607.21017) TableVerse-100K is an unprecedented, massive-scale Real2Sim tabletop dataset constructed directly from unstructured, in-the-wild single internet frames. By replacing traditional probabilistic hallucinations with a deterministic perception-to-physics workflow (featuring Layout-Consistent Collision Rectification, LCCR), TableVerse bridges the reality gap to deliver physically consistent, simulation-ready digital twins. The dataset comprises **100K unique tabletop environments**, nearly **1M distinct object instances** spanning over **35K semantic categories**, completely annotated with multi-dimensional physical properties and paired with continuous, collision-free expert manipulation trajectories. > ⚠️ **Data Limitation & Disclaimer:** Please note that because TableVerse-100K is generated through a fully automated, end-to-end Real-to-Sim pipeline to achieve unprecedented scale, the dataset is inherently not perfect and contains geometric artifacts or semantic misalignments in some long-tail scenarios. --- ## 📁 Repository Structure ```text TableVerse/ ├── textures/ ├── scenes/ │ ├── <scene_uid>/ │ │ ├── assets/ │ │ ├── textures/ │ │ ├── scene.glb │ │ ├── scene.xml │ │ ├── scene_with_plane.xml │ │ └── relabel_name.json │ └── ...