RoboInter-Data Dense Intermediate-Representation Corpus

基于DROID与RH20T真实机器人数据构建的密集逐帧中间表征语料；固定修订公开非gated，643项合计约407.79GB，抽样质量统计文件确认235,920段。数据标准license字段缺失，隐藏门控文本与README的上游许可指引冲突，必须逐源核验。

开放状态: 数据可下载·许可证冲突
许可证: No standard license field; hidden gate text says CC BY-NC-SA 4.0; README defers to RoboInter, DROID and RH20T licenses
商业使用: not established for the full mixture; RH20T contains noncommercial scenes
人工评审结论: verified_public_nongated_fixed_revision_407786031053_bytes_license_conflict
机器人: Franka Research 3 / DROID setup、RH20T multi-arm configurations
任务: robot manipulation, intermediate-representation understanding, VQA, VLA planning and action prediction
模态: primary RGB video、wrist RGB video、action、end-effector state、language instruction、subtask、primitive skill、segmentation、object/gripper/affordance boxes、placement proposal、future trace、contact frame/point、quality rating
格式: LeRobot v2.1、Parquet、MP4、LMDB、PKL、NPZ、JSON/JSONL
采集/生成方式: dense automatic/curated annotation and LeRobot repackaging of DROID and RH20T physical robot episodes
已知问题: The 235,920 units are derived multi-view/annotation episodes rather than wholly new physical collections; full 407.79GB was not downloaded; Hugging Face viewer shows 7,246 validation rows; license metadata conflicts and upstream RH20T scene-level commercial restrictions require audit; July VLA/World code is not released.