USDCraft: Geometrically Grounded Programmatic Modeling of Articulated 3D Assets for Simulation
USDCraft uses an LLM to write simulation-ready articulated 3D assets from partial mesh evidence.
The paper frames the problem around real-to-sim robot manipulation, where corrupted or incomplete object meshes can break mesh-based articulation methods. USDCraft converts the observed geometry into a metric text description, has the model generate executable asset programs, then rechecks candidates against the same representation to guide edits. The output is articulated USD assets with physical properties that load into Isaac Sim without manual adjustment. The authors report leading articulation recovery on two benchmarks and validation in real-to-sim-to-real manipulation. HF Daily Papers' note
The paper frames the problem around real-to-sim robot manipulation, where corrupted or incomplete object meshes can break mesh-based articulation methods. USDCraft converts the observed geometry into a metric text description, has the model generate executable asset programs, then rechecks candidates against the same representation to guide edits. The output is articulated USD assets with physical properties that load into Isaac Sim without manual adjustment. The authors report leading articulation recovery on two benchmarks and validation in real-to-sim-to-real manipulation. HF Daily Papers' note
score 5