CoDimRecon Targets Sim-Ready Deformable 3D Scenes

CoDimRecon’s strongest idea is treating physical behavior as part of 3D reconstruction, not an afterthought. The approach is promising for robotics and simulation pipelines, but remains research-stage rather than a turnkey asset-generation tool.

• –Dimension-specific representations make ropes, sheets, and cushions more practical to simulate than generic meshes.
• –Agent-guided behavioral tests can expose failures that visual or numerical checks miss, such as incorrect paper-folding behavior.
• –The system combines scene-level layout, generated object meshes, editable geometry, articulation, and material modeling in one pipeline.
• –Results on Replica and ScanNet++ show competitive reconstruction accuracy, while deformable examples are capability demonstrations rather than cross-method benchmarks.
• –The inferred physical parameters are effective simulation values, not verified recovery of real-world material properties. [Project page](https://shuzhaoxie.github.io/CoDimRecon/)