STABLE

Simulation-Ready Tabletop Layout Generation via a Semantics-Physics Dual System

Zhen Luo1,2,*, Yixuan Yang2,3,*, Xudong Xu3,†, Jinkun Hao4, Zhaoyang Lyu3, Feng Zheng2,6,†, Jiangmiao Pang3, Yanwei Fu1,5
SII1, SUSTech2, Shanghai AI Laboratory3, SJTU4, FDU5, Spatialtemporal AI6
*Equal contribution. †Corresponding authors.
ICML 2026
Paper arXiv Code

TL;DR: STABLE couples an LLM-based Semantic Reasoner with a geometry-aware Physics Corrector to generate task-aligned, collision-free tabletop scenes that are ready for simulation and robot manipulation.

STABLE overview for tabletop scene generation

OVERVIEW VIDEO

From task instructions to simulation-ready scenes

Robot Operation Videos

Generated scenes for manipulation rollouts.

Scene 01 Video slot reserved

Place the first manipulation video at assets/videos/robot-scene-1.mp4.

Generated Scene 01

Scene 02 Video slot reserved

Place the second manipulation video at assets/videos/robot-scene-2.mp4.

Generated Scene 02

Visualization Results

Task-aligned layouts with fewer physical failures.

Qualitative comparisons highlight common baseline failures, including interpenetration, missing task objects, and broken spatial relations. STABLE preserves task semantics while producing simulation-ready tabletop geometry.

Qualitative comparison with baseline methods

Applications

Editing and rearrangement.

Incremental scene editing examples
Incremental editing preserves scene semantics while modifying local layout structure.
Tabletop rearrangement examples
The Physics Corrector recovers physically consistent tabletop rearrangements.

Method

A progressive semantics-physics loop.

STABLE alternates semantic scene construction with physics-aware pose correction, expanding from task-critical objects to contextual background objects while keeping each intermediate layout feasible.

STABLE method overview

Citation

BibTeX

@inproceedings{luo2026stable,
  title = {STABLE: Simulation-Ready Tabletop Layout Generation via a Semantics-Physics Dual System},
  author = {Luo, Zhen and Yang, Yixuan and Xu, Xudong and Hao, Jinkun and
            Lyu, Zhaoyang and Zheng, Feng and Pang, Jiangmiao and Fu, Yanwei},
  booktitle = {International Conference on Machine Learning},
  year = {2026}
}