HiFi-UMI-2K High-Fidelity Robot-Free Bimanual Manipulation Dataset

HiFi-UMI-2K High-Fidelity Robot-Free Bimanual Manipulation Dataset

HiFi-UMI公开的2,000小时机器人无关双手操作数据，固定Hugging Face revision约16.04TB、5,176个文件，采用六路同步视频、双手末端轨迹、夹爪状态、语言、任务与质量控制字段。采集系统报告约3mm局部末端精度和低于40微秒跨传感器偏移；公开包为20,000+小时源语料的子集，不等同论文4,000小时预训练混合。

开放状态

已核验开放

许可证

CC-BY-4.0

商业使用

permitted by CC-BY-4.0 with attribution; downstream model and third-party software terms remain

人工评审结论

verified_public_nongated_fixed_revision_large_scale_real_human_manipulation_dataset

机器人

robot-free HiFi-UMI bimanual capture system、downstream bimanual deployment robot

任务

large-scale daily bimanual manipulation with language and task metadata; the fixed file tree is sharded rather than a short closed task list

模态

stereo head RGB、four hand fisheye RGB streams、bimanual end-effector pose、gripper state、language annotation、task metadata、validity and quality-control masks、normalization statistics

格式

LeRobot v3-style sharded Parquet + MP4 with meta/info.json, modality, stats, tasks and episode tables

采集/生成方式

robot-free natural human manipulation with head-mounted offline stereo-inertial SLAM, common-frame hand markers, encoder-equipped grippers and shared GPIO hardware trigger

已知问题

Full 16TB bytes were not downloaded; capture hardware/software is not released as a complete reproducible kit; 2K public release, 4K pretraining mixture, 20K+ source hours and 4.32M+ source episodes must remain distinct; no independent training report.

论文原文 · 官方项目页 · 官方数据固定revision