A new arXiv preprint introduces GRASE, a system for real-to-sim conversion that takes video of a real environment and produces an editable 3D scene. The stated goal is to preserve the geometry and layout of what was observed, which is a core requirement for downstream simulation and editing tasks.

The abstract contrasts GRASE with existing methods, which it characterizes as following a one-pass, loosely coupled paradigm. GRASE is described as geometry-native, suggesting a tighter integration between geometric understanding and scene modeling. However, the abstract is truncated, so specific technical details and experimental results are not available in the source.

This summary is based solely on the single arXiv source; there are no other sources to compare or contrast. Further details would require the full paper or additional reporting.