Towards Physically Grounded Articulated 3D Assets: A Multi-Agent Pipeline and Benchmark
Artisan adds simulation feedback to articulated 3D asset generation so outputs are judged on whether they can move, hold together, and behave physically.
The paper frames the task as moving beyond visual reconstruction toward functional 3D assets with plausible articulation, collision-free motion, gravity stability, and mass/material attributes. It introduces a multi-agent pipeline built on Articraft and a benchmark covering spatial fidelity, physical plausibility, downstream task performance, and overall quality. The benchmark uses 100 reference assets, 50 stability prompts, and 50 ABO objects with reference masses. In tests across more than 10 models, Artisan is reported to keep competitive spatial fidelity while producing the lowest collision rates and highest stability rates. HF Daily Papers' note
The paper frames the task as moving beyond visual reconstruction toward functional 3D assets with plausible articulation, collision-free motion, gravity stability, and mass/material attributes. It introduces a multi-agent pipeline built on Articraft and a benchmark covering spatial fidelity, physical plausibility, downstream task performance, and overall quality. The benchmark uses 100 reference assets, 50 stability prompts, and 50 ABO objects with reference masses. In tests across more than 10 models, Artisan is reported to keep competitive spatial fidelity while producing the lowest collision rates and highest stability rates. HF Daily Papers' note
score 4