HandEdit Egocentric Human-to-Robot Dexterous-Hand Visual Data

HandEdit Egocentric Human-to-Robot Dexterous-Hand Visual Data

HandEdit-Full 是从五个第一视角人手物数据源经手部移除、背景修复、运动重定向、IK/碰撞筛选、渲染和人工筛选生成的机器人手/手臂视觉伪 GT 数据。论文称其覆盖超过2亿图像级编辑实例、30万以上clips、600+场景、400+任务、1,100+物体和26个目标URDF；固定 Hugging Face 版本公开、非 gated，页面显示5.51TB和97,526文件。它不是机器人动作、触觉、力或真机控制轨迹，且数据为非商用、禁止演绎许可。

开放状态

已核验开放

许可证

CC-BY-NC-ND-4.0 declared on the fixed Hugging Face dataset revision

商业使用

prohibited by CC-BY-NC-ND-4.0; no-derivatives restriction applies, and upstream human-data rights remain unverified

人工评审结论

fixed_revision_public_nongated_551tb_visual_synthetic_dataset_noncommercial_no_derivatives

机器人

26 target dexterous-hand and hand-arm URDF embodiments

任务

egocentric human-hand to robot-hand image editing, cross-embodiment dexterous visual generation and benchmark evaluation

模态

egocentric RGB、robot-hand edited RGB、URDF-conditioned rendered pseudo-GT、hand masks、benchmark metadata

格式

Hugging Face dataset repository; file viewer timed out and README was empty at review, so complete schema is not yet verified

采集/生成方式

five upstream egocentric human hand-object sources processed through segmentation, inpainting, motion retargeting, IK, collision filtering, rendering and manual selection

已知问题

No full download, file hashing, schema audit, upstream privacy/right audit or independent reproduction completed. The HF viewer timed out and README was empty. Data is visual pseudo-GT, not action/tactile/force data; commercial use and derivatives are prohibited.

论文原文 · 官方项目页 · 官方评测代码固定提交 · 官方数据固定版本