HandEdit: A Unified Benchmark for Egocentric Human-to-Robot Dexterous Hand Image Editing

Zhenjie Yang^2,3*, Xingyu Jiao^1*, Guopeng Zhong³, Shuzhe Yang³, Shi Che³, Chao Wu¹, Chenyu Jiang¹,
Dongjie Zhang¹, Yideng Zhang³, Zheng Zhang³, Muyun Jiang⁵, Haisheng Su³, Shuang Jin², Donghang Zhang²,
Chao Yang⁶, Li Chen⁴, Hongyang Li⁴, Zuxuan Wu¹, Yu-Gang Jiang¹, Xiaosong Jia^1†, Junchi Yan^3†

¹ Fudan University ² Inspire Robots ³ Shanghai Jiao Tong University
⁴ The University of Hong Kong ⁵ Nanyang Technological University ⁶ Shanghai AI Laboratory
^* Equal contribution ^† Corresponding authors
Contact: yangzhenjie@sjtu.edu.cn, jiaxiaosong@fudan.edu.cn

Paper Code Dataset Results BibTeX

Abstract

Robotic manipulation with dexterous hands is a cornerstone of Embodied AI, yet embodiment-aware teleoperation data remains costly to collect. Egocentric human videos offer a scalable alternative, but human and robotic hands differ profoundly in appearance, articulation, kinematics, and camera-relative geometry. HandEdit formulates this gap as an embodiment-aware image-editing problem. The benchmark asks an editor to replace the visible human hand or hand-arm region with a requested dexterous robot embodiment, while preserving object state, task semantics, contact relationships, viewpoint, and surrounding scene structure.

Embodiment Transfer

Apple cutting

HandEdit Dataset

EgoDex90M frames

338K clips · 500 household objects · 194 tasks

ARCTIC2.1M frames

Articulated objects · tool use · grasping

OakInk24.01M frames

75 objects · 150 tasks

HOI4D2.4M frames

Room-scale diversity · 54 tasks

HO-Cap656K frames

Pick-and-place · handover · tool use

Robot Embodiments

Benchmark Results

Model	Access	LPIPS ROI ↓	FID ROI ↓	Removal ↑	Struct ↑	ID ↑	Interaction ↑	VLM ↑

Radar plots comparing image editors on Hand-only and Hand-Arm metrics

Conclusion

GPT-Image-2 provides the strongest overall baseline.

VLM-based judgment is useful but not sufficient.

Perceptual quality alone does not guarantee editing task success.

The main challenge lies in embodiment-aware editing.

Qualitative Results

Qualitative comparison across commercial and open-source editors

BibTeX

@article{yang2026handedit,
    title={{HandEdit}: A Unified Benchmark for Egocentric Human-to-Robot Dexterous Hand Image Editing}, 
    author={Zhenjie Yang and Xingyu Jiao and Guopeng Zhong and Shuzhe Yang and Shi Che and Chao Wu and Chenyu Jiang and Dongjie Zhang and Yideng Zhang and Zheng Zhang and Muyun Jiang and Haisheng Su and Shuang Jin and Donghang Zhang and Chao Yang and Li Chen and Hongyang Li and Zuxuan Wu and Yu-Gang Jiang and Xiaosong Jia and Junchi Yan},
    year={2026},
    eprint={2608.12122},
    archivePrefix={arXiv},
    primaryClass={cs.RO}
}