Real-World Bimanual Dexterous Grasp Dataset for Unitree H1-2

Unitree H1-2双臂、两只Inspire RH56DFTP灵巧手的真实大物体协同抓取数据。论文记录33种物体、每物体6至20种姿态、共353次15Hz抓取序列；模态含RGB、深度、指令/实际关节角、关节扭矩和分布式触觉。官方仓固定commit含约251MB代码与样例数据，README另给出约40GB完整百度网盘；根目录没有LICENSE，数据仅声明非商业研究，因此属于可下载但许可不完整。

开放状态: 样例与完整包入口已核验·许可证不完整
许可证: Non-commercial research only stated in README; no standalone dataset license
商业使用: not permitted by README statement; exact terms incomplete
人工评审结论: sample_and_full_archive_links_verified_license_incomplete
机器人: Unitree H1-2、two Inspire RH56DFTP dexterous hands
任务: cooperative bimanual grasping、large-object lifting、stable holding、placing
模态: head RGB、head depth、segmented point cloud、commanded joint angles、actual joint angles、joint torque、distributed tactile、object masks、selected grasp frame
格式: repository samples include JPG, PNG, PLY, NPY, CSV and JSON; complete archive format not independently enumerated
采集/生成方式: Apple Vision Pro OpenXR hand/wrist tracking retargeted to Unitree H1-2 dual arms and Inspire hands; RealSense D435i head RGB-D
已知问题: No root LICENSE, no standalone dataset terms, no full-archive checksum, no model weights, constant lighting/background, one humanoid embodiment, and no independent reproduction. GitHub sample is not the full 40GB corpus.

论文原文 · 官方代码与样例数据 · 官方仓声明的完整数据包