Reproduce “TableVerse: A Large-scale Tabletop Dataset with Real-world Grounded Layouts for Generalizable Manipulation” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Boyuan Wang, Yue Zhang, Xutao Xue, Xueyu Song, Yu Sun
TableVerse presents an automated Real2Sim pipeline that transforms unstructured internet images into 100,000 high-fidelity, physically consistent, simulation-ready tabletop environments for robotic manipulation research. This system achieved a 0.0% scene collision rate and superior MLLM-evaluated layout and geometry fidelity compared to existing 3D scene reconstruction methods.
Key insights
- Deterministic reconstruction from real-world observations offers a more effective path to generating physically accurate and manipulation-ready scene layouts than imaginative or procedural methods.
- Combining geometric collision rectification with physics-based stabilization is crucial for transforming raw reconstructed scenes into stable, simulation-ready environments.
- Multi-modal large language models (MLLMs) can be integrated into an automated workflow for scalable, high-quality data curation, annotation, and task generation, including complex spatial relation predicates.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.