TableVerse: A Large-scale Tabletop Dataset with Real-world Grounded Layouts for Generalizable Manipulation

Researchers introduced TableVerse, a fully automated Real2Sim pipeline for robotic manipulation data generation.

It reconstructs real-world tabletop scenes from unstructured images and augments them with synthetic pick-and-place trajectories.

The result is TableVerse-100K, a 100,000-scene dataset with physically plausible, collision-free demonstrations.

The dataset offers a scalable source of realistic scenes that may improve robot policies’ generalization in cluttered environments.