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

TableVerse introduces an automated Real2Sim pipeline that reconstructs tabletop environments from unstructured, in-the-wild internet media rather than hallucinating layouts or relying on simplified procedural generation. The reconstructed scenes are intended to preserve metric scale, authentic object topology, and verified mechanical stability. The pipeline also generates task-conditioned, collision-free pick-and-place demonstrations. Its resulting TableVerse-100K corpus contains 100,000 unique, physically consistent environments paired with interactive manipulation trajectories. The paper positions this dataset as a scalable source of realistic training data for generalizable robotic manipulation, although the supplied abstract does not report benchmark results, sim-to-real evaluations, or detailed reconstruction accuracy.