Data scale
2.6M cluttered scenes and 0.4B validated grasp poses.
Scroll to explore ↓
01
Abstract
Data diversity
Five scene regimes, diverse supports, objects, and views.
OmniDex model
Multimodal generation with Soft-WTA and physical constraints.
02
OmniDex Benchmark
Data Generation
Reusable seed grasps are validated in dynamically settled cluttered scenes.
OmniDex Model
Multimodal grasp generation with physical constraints and ranking.
03
Results
• Five benchmark regimes
• Standard tabletop
• Mixed supports
• Confined boxes
• Structured grids
• Retail shelves
04
BibTeX
@inproceedings{fang2026omnidex,
title={OmniDex: Scaling Dexterous Hand Grasping to Diverse Cluttered Scenes},
author={Fang, Naiyu and Luo, Zhongjin and Mo, Yuxin and Huang, Siyuan and Liu, Jianbo and Liu, Yufei and Zhou, Zheyuan and Jin, Chenkai and Wang, Xiaogang and Li, Hongsheng},
booktitle={Advances in Neural Information Processing Systems},
year={2026}
}