--- license: other task_categories: - robotics tags: - robotics - dexterous-manipulation - teleoperation - motion-capture - human-hand - imitation-learning - hdf5 size_categories: - 10GB --- # CAP —— 杯序整理任务(右手)灵巧操作数据集 CAP(Capture Dataset)是一套**人类手部动作采集 + 灵巧操作演示**数据集,记录了操作者用右手完成「**按序摆放杯子(cups_order)**」任务的全过程。数据以 DexCap 风格的采集管线记录,包含**压缩后的轨迹 HDF5** 与**多视角相机的原始图像帧**,可直接用于动作复现、模仿学习与手部姿态分析。 - 仓库:`dex-robot/cap` - 主文件:`cap.zip`(10,697,737,890 字节 ≈ 9.96 GiB) - 收录内容:18,939 个文件 + 24 个目录条目 - SHA256:`f91c11ee1761822fa875d4783c1fd9388f6fa3f0fa3feeef6cd5c20775220c10` - 采集日期:2026-07-07 - 任务:cups_order(按序摆放杯子) - 手别:右手(right hand) --- ## 1. 数据总览 | 项目 | 内容 | | --- | --- | | 任务 | cups_order(按序摆放杯子) | | 采集手别 | 右手 | | 采集日期 | 2026-07-07 | | 有效采集次数 | 5 次(16:46:34 / 16:54:35 / 16:56:54 / 16:59:04 / 17:01:24) | | 轨迹文件 | 5 个 `.hdf5`(每次采集一个) | | 相机视角 | 6 路(左手 / 右手 / 头部双目 / 头部全局 RGB / 头部全局深度) | | 图像帧 | 18,914 张 `.png` + 5 段 `.mp4` | | 其他 | 3 个 `.json`、6 个 `.txt`、5 个 `.npy`、1 个 `.raw` | | 压缩包大小 | 10.7 GB(10,697,737,890 字节) | --- ## 2. 目录结构 ``` cap/ ├── hdf5/ # 每次采集一条压缩轨迹 │ ├── cups_order_right_hand_20260707_164634/ │ │ └── cups_order_right_hand_20260707_164634.hdf5 │ ├── cups_order_right_hand_20260707_165435/ │ │ └── cups_order_right_hand_20260707_165435.hdf5 │ ├── cups_order_right_hand_20260707_165654/ │ │ └── cups_order_right_hand_20260707_165654.hdf5 │ ├── cups_order_right_hand_20260707_165904/ │ │ └── cups_order_right_hand_20260707_165904.hdf5 │ └── cups_order_right_hand_20260707_170124/ │ └── cups_order_right_hand_20260707_170124.hdf5 └── raw_data/ # 原始多模态记录(图像 + 视频 + 标定) └── cups_order_right_hand_20260707_164634/ └── cups_order_right_hand_20260707_164634_extracted/ └── observe/ └── vision/ ├── left_hand_rgb/ # 左手腕视角(video + metadata) ├── right_hand_rgb/ # 右手腕视角 ├── head_zed_left/ # 头部 ZED 左目 ├── head_zed_right/ # 头部 ZED 右目 ├── head_global_rgb/ # 头部全局 RGB ├── head_global_depth/ # 头部全局深度 └── *_frames/ # 各视角抽帧后的 PNG 序列 ``` --- ## 3. 采集会话清单 | 会话 ID | 起始时间 | HDF5 轨迹 | 原始帧 | | --- | --- | --- | --- | | `cups_order_right_hand_20260707_164634` | 16:46:34 | ✅ | ✅ | | `cups_order_right_hand_20260707_165435` | 16:54:35 | ✅ | — | | `cups_order_right_hand_20260707_165654` | 16:56:54 | ✅ | — | | `cups_order_right_hand_20260707_165904` | 16:59:04 | ✅ | — | | `cups_order_right_hand_20260707_170124` | 17:01:24 | ✅ | — | > 说明:完整的多视角原始帧(18,915 张 PNG + 5 段 MP4)随**首个会话** `164634` 一同发布;其余 4 次采集提供压缩后的 HDF5 轨迹。 ## 4. 相机视角与帧数 | 视角目录 | 帧数(PNG) | 视频 | | --- | --- | --- | | `head_global_depth_frames` | 5,565 | ✅ | | `right_hand_rgb_frames` | 3,089 | ✅ | | `head_zed_right_frames` | 2,629 | ✅ | | `left_hand_rgb_frames` | 2,563 | ✅ | | `head_zed_left_frames` | 2,546 | ✅ | | `head_global_rgb_frames` | 2,522 | ✅ | 帧文件命名规则:`frame_%06d.png`。**编号并非严格连续**(采集过程中存在丢帧),使用时请按文件名排序取序列,不要假设 `i` 与 `i+1` 相邻。 ### 数据预览 ![六视角总览](preview_overview.png) ![六视角同步动图](preview_grid.gif) | 预览文件 | 内容 | | --- | --- | | `preview_overview.png` | 6 路视角单帧总览 | | `preview_grid.gif` | 6 路视角同步动图(24 帧) | | `preview_right_hand_rgb.gif` | 右手腕视角动态预览 | | `preview_head_global_rgb.gif` | 头部全局 RGB 动态预览 | | `preview_head_global_depth.gif` | 头部全局深度动态预览 | --- ## 5. 下载 ### 5.1 huggingface-cli(推荐,支持断点续传) ```bash pip install -U "huggingface_hub[cli]" export HF_ENDPOINT=https://hf-mirror.com # 国内网络可加,境外可省略 huggingface-cli download dex-robot/cap cap.zip --repo-type dataset --local-dir ./cap_data ``` ### 5.2 Python ```python from huggingface_hub import hf_hub_download p = hf_hub_download("dex-robot/cap", "cap.zip", repo_type="dataset") print(p) ``` ### 5.3 直链 ``` https://huggingface.co/datasets/dex-robot/cap/resolve/main/cap.zip ``` --- ## 6. 解压与校验 ```bash # 1) 大小核对 stat -c %s cap.zip # 应为 10697737890 # 2) SHA256 核对 sha256sum cap.zip # f91c11ee1761822fa875d4783c1fd9388f6fa3f0fa3feeef6cd5c20775220c10 # 3) 逐文件 CRC 校验(18963 项全部 OK 即为通过) unzip -t cap.zip # 4) 解压到当前目录,得到 cap/ unzip -q cap.zip du -sh cap ``` 预期结果: ``` cap/ ├── hdf5/ (5 个 .hdf5) └── raw_data/ (18,934 个文件) ``` --- ## 7. 读取示例 ### 7.1 读取 HDF5 轨迹 ```python import h5py path = "cap/hdf5/cups_order_right_hand_20260707_164634/cups_order_right_hand_20260707_164634.hdf5" with h5py.File(path, "r") as f: def show(name, obj): if isinstance(obj, h5py.Dataset): print(f"D {name:50s} shape={obj.shape} dtype={obj.dtype}") else: print(f"G {name}") f.visititems(show) ``` ### 7.2 读取图像帧 ```python import cv2, glob frames = sorted(glob.glob( "cap/raw_data/cups_order_right_hand_20260707_164634/" "cups_order_right_hand_20260707_164634_extracted/observe/vision/" "head_global_rgb_frames/frame_*.png")) print("帧数:", len(frames)) img = cv2.imread(frames[0]) print(img.shape) # (H, W, 3) ``` --- ## 8. 推荐环境 ```bash pip install h5py opencv-python numpy ``` - Python ≥ 3.8 - 解压后约需 15 GB 可用磁盘(原始 PNG 序列体积较大) - 读取深度数据时注意 `head_global_depth` 可能为 16-bit PNG 或伴随 `.npy` / `.raw` 原始深度 --- ## 9. 许可与引用 - 许可:见仓库 `LICENSE`(未单独指定时请联系数据发布方) - 引用本数据集时请注明来源: ```bibtex @misc{cap_dataset_2026, title = {CAP: Cups Ordering Dexterous Manipulation Capture Dataset (Right Hand)}, author = {dex-robot}, year = {2026}, howpublished = {\url{https://huggingface.co/datasets/dex-robot/cap}} } ``` --- ## 10. 常见问题 **Q:下载中断怎么办?** `huggingface-cli download` 支持断点续传,重新执行同一条命令即可继续;也可加 `--max-workers 1` 降低并发提升稳定性。 **Q:`unzip` 报 "End-of-central-directory signature not found"?** 说明下载不完整、文件被截断,请校验文件大小是否为 **10,697,737,890 字节**、SHA256 是否匹配,不一致请重新下载。 **Q:只想用轨迹,不想下 10 GB 图像?** 目前仅提供整包发布;有细分需求请在仓库 Discussion 中提出。 **Q:深度图和 RGB 怎么对齐?** 同视角的 `*_frames` 与 MP4 按帧序号(`frame_%06d`)一一对应;跨视角对齐请使用 `raw_data/.../observe/vision/` 下的标定 JSON。 --- ## 11. 已知限制 1. **帧编号不连续**:各视角 `frame_%06d.png` 的编号存在跳号(采集过程中的丢帧),例如 `right_hand_rgb_frames` 实际 3,089 帧、最大编号 3,746。请按文件名排序后的顺序使用,不要通过 `i+1` 推断下一帧。 2. **原始帧仅覆盖首个会话**:5 次采集中,只有 `164634` 附带完整的多视角 PNG / MP4;其余 4 次仅有 HDF5 轨迹。 3. **HDF5 内部结构随版本演进**:不同采集时间生成的 HDF5 其 group / dataset 名称可能有差异,读取时请先 `visititems` 打印结构再取数。 4. 本数据集由一次中断的打包过程抢救重建而来(已通过全部 18,963 项 CRC 校验),末尾 1 个不完整的 PNG 帧已被剔除。