ARMO: Autoregressive Rigging for Multi-Category Objects


Mingze Sun* 1   Shiwei Mao* 1   Keyi Chen1   Yurun Chen1   Shunlin Lu2   Jingbo Wang3   Junting Dong† 3   Ruqi Huang† 1  

1 Tsinghua Shenzhen International Graduate School, China  2 The Chinese University of Hong Kong, Shenzhen  3 Shanghai AI Laboratory, China
* Indicates Equal Contribution      † Indicates Corresponding Author

ARMO has been accepted by ICCV 2025! See you in Hawaii!
Paper arXiv Video (coming soon) Code (coming soon) Dataset (coming soon)
Demo (coming soon)

We propose a novel rigging framework, ARMO, an autoregressive model that simultaneously predicts both joint positions and connectivity relationships. For rigging 3D shapes, ARMO can manifest significant diversity in style, pose, and category. To train ARMO, we propose OmniRig, a large-scale rigging dataset, which covers diverse object categories with detailed rigging annotations.



Brief Introduction of OmniRig

We introduce OmniRig, a comprehensive and large-scale dataset with detailed rigging annotations. Our dataset is constructed from three key sources: ModelResource, ObjaverseXL, and publicly available free data collected from the internet. In total, OmniRig comprises 79,499 meshes, each accompanied by detailed rigging information.



Comparison with Challenging Methods

We compare our method with other skeleton generation methods on OmniRig. Our method can generate reasonable skeleton results for diverse object categories and inputs with complex poses.



Demo

BibTeX

@article{sun2025armo,
  title={ARMO: Autoregressive Rigging for Multi-Category Objects},
  author={Sun, Mingze and Mao, Shiwei and Chen, Keyi and Chen, Yurun and Lu, Shunlin and Wang, Jingbo and Dong, Junting and Huang, Ruqi},
  journal={arXiv preprint arXiv:2503.20663},
  year={2025}
}