# SimuScene: Simulation-Ready Compositional 3D Scene Reconstruction from a Single Image

Source: [SimuScene: Simulation-Ready Compositional 3D Scene Reconstruction from a Single Image](<https://www.sota2.com/research/paper/6a210696c10484bac23f0233>)

Collection index: [llms.txt](<https://www.sota2.com/research/llms.txt>)

Data updated: 2026-06-04T05:01:01.134Z

Authors: Inhee Lee, Sangwon Baik, Sungjoo Kim, Hyeonwoo Kim, Hyunsoo Cha, Hanbyul Joo

Published: 2026-06-02T17:59:59.000Z

## Abstract

Reconstructing interactive, simulation-ready 3D scenes from a single image is a critical bottleneck for robotic manipulation. While recent single-image lifters recover plausible per-object shapes, composing them yields scenes that collapse under physical simulation due to interpenetrating, hovering, or sinking objects. Existing physics-aware methods address this strictly as a post-hoc layout correction, leaving the underlying geometric errors unresolved. To address this, we introduce SimuScene, a compositional 3D reconstruction pipeline that puts physics in the loop of shape and layout estimation. Rather than using physics merely for layout cleanup, we utilize the physics engine as a diagnostic measurement tool during the generative process itself. By diagnostically simulating reconstructed objects under gravity, we convert penetration and support failures into quantitative correction signals that drive gravity-axis stretching and amodal shape resampling. This physics-informed feedback loop mitigates accumulated reconstruction errors and produces a stable, simulation-ready compositional 3D scene. Extensive experiments demonstrate state-of-the-art performance on physical stability and geometric alignment benchmarks. We further highlight SimuScene's utility by deploying reconstructed environments in humanoid control and robot-arm manipulation tasks.

## Recorded benchmark results

A research result applies to its stated dataset, split, metric and evaluation configuration. Compare like-for-like settings, preserve the metric's unit and direction, and distinguish paper-reported results from independent reproduction. The best result recorded here is not a claim of exhaustive, current global coverage.

- [3D Scene Reconstruction on GraspClutter6D YCB-V scene split (test)](<https://www.sota2.com/research/sota/3d-scene-reconstruction-on-graspclutter6d-ycb-v-scene-split-test.md>): 3D Scene Reconstruction; GraspClutter6D YCB-V scene split (test); ABO (fh/fo): 0.248.

- [3D Scene Reconstruction on AriaDigitalTwin](<https://www.sota2.com/research/sota/3d-scene-reconstruction-on-ariadigitaltwin.md>): 3D Scene Reconstruction; AriaDigitalTwin; 3D ABO (fh/fo): 29.3.

- [Instance Mask Extraction on GraspClutter6D](<https://www.sota2.com/research/sota/instance-mask-extraction-on-graspclutter6d.md>): Instance Mask Extraction; GraspClutter6D; IoU (fh/fo): 21.3.

- [Instance Mask Extraction on AriaDigitalTwin](<https://www.sota2.com/research/sota/instance-mask-extraction-on-ariadigitaltwin.md>): Instance Mask Extraction; AriaDigitalTwin; IoU (fh/fo): 23.4.

- [Instance Mask Extraction on GenWild](<https://www.sota2.com/research/sota/instance-mask-extraction-on-genwild.md>): Instance Mask Extraction; GenWild; IoU (fh/fo): 49.4.

- [3D Scene Reconstruction on GenWild](<https://www.sota2.com/research/sota/3d-scene-reconstruction-on-genwild.md>): 3D Scene Reconstruction; GenWild; ABO (fh/fo): 0.507.

## Original sources and implementation

[Original paper](<https://arxiv.org/abs/2606.03994>)

[PDF](<https://arxiv.org/pdf/2606.03994.pdf>)