UniArt: Generating 3D articulated objects with open-set articulation beyond retrieval

ICLR 2026 Conference Submission25187 Authors

Keywords: 3d Generation, Embodied AI

Abstract: Articulated objects are central in the field of realistic simulation and robot learning, enabling dynamic interactions and task-oriented manipulation. However, manually annotating these objects is labor-intensive, motivating the need for automated generation solutions. Previous methods usually rely on retrieving part structures from existing datasets, which inherently restricts diversity and causes geometric misalignment. To tackle these challenges, we present UniArt, an end-to-end framework that directly synthesizes 3D meshes and articulation parameters in a unified manner. We decompose the problem into three correlated tasks: geometry generation, part segmentation, and articulation prediction, and then integrate them into a single diffusion-based architecture. By formulating both part segmentation and joint parameter inference as open-set problems, our approach incorporates open-world knowledge to generalize beyond training categories. We further enhance training with a large-scale, enriched dataset built from PartNet-Mobility, featuring expanded part and material diversity. Extensive evaluations show that UniArt substantially outperforms existing retrieval-based methods in mesh quality and articulation accuracy, especially under open-set conditions. Code will be publicly available to foster future research in the 3D generation and robotics societies.

Primary Area: applications to robotics, autonomy, planning

Submission Number: 25187