Chinese_Commercial_Kitchen_Manipulation_Dataset_Preview
收藏资源简介:
Chinese Commercial Kitchen Manipulation Dataset — Sample Pack v0.1 是亚洲首个真实商业厨房操作数据集,专为机器人学、具身人工智能、模仿学习和视觉动作理解研究设计。该数据集在中国广东省中山市一家运营中的商业中餐厅厨房采集,由拥有约20年经验的专业厨师演示真实烹饪工作流程,而非模拟演示。数据内容包含多视角同步或任务对齐的视频采集:第一人称视角(头戴式运动相机)、侧视角(固定手机相机)和俯视角(Intel RealSense D435I RGB-D相机)。样本包当前包含两个烹饪任务:切菜任务提供三个相机视角(第一人称、侧视、俯视)以及对齐的深度数据(HDF5格式,480×848分辨率,2379帧);翻炒任务提供两个视角(第一人称和侧视)。深度数据中,像素值为65.535米表示无效深度测量,典型有效像素率约为86%。技术规格方面:样本包中的第一人称视频为1080p预览编码(1920×1080,30fps),完整4K源录像(3840×2160,29.97fps)可申请获取;侧视角视频为1920×1080分辨率(60fps或30fps);俯视角RGB视频为1280×720分辨率,15fps;深度数据为480×848分辨率,浮点32位,单位为米。数据集适用于机器人操作、模仿学习、动作理解、深度估计等任务,特别关注工具使用、双手协调和精细食物操作等挑战。完整数据集计划扩展至包括包饺子、揉面、炸、煎、炖煮、勾芡、腌制/调味等7个额外烹饪任务,并可提供任务级或动作级标注。数据采集已获得所有参与者的充分知情同意。
Chinese Commercial Kitchen Manipulation Dataset — Sample Pack v0.1 is Asia's first real-world commercial kitchen manipulation dataset, specifically designed for research in robotics, embodied artificial intelligence, imitation learning, and visual action understanding. The dataset was collected in an operating commercial Chinese restaurant kitchen in Zhongshan City, Guangdong Province, China, with professional chefs possessing approximately 20 years of experience demonstrating authentic cooking workflows rather than simulated ones. The data includes multi-view synchronized or task-aligned video recordings: first-person view (captured via a head-mounted action camera), side view (captured via a fixed smartphone camera), and top-down view (captured via an Intel RealSense D435I RGB-D camera). The current sample pack contains two cooking tasks: the vegetable cutting task provides three camera views (first-person, side view, top-down view) along with aligned depth data in HDF5 format (480×848 resolution, 2379 frames); the stir-frying task provides two views (first-person and side view). For the depth data, a pixel value of 65.535 meters indicates an invalid depth measurement, with a typical valid pixel rate of approximately 86%. Regarding technical specifications: the first-person videos in the sample pack are encoded as 1080p previews (1920×1080, 30fps), while the full 4K source footage (3840×2160, 29.97fps) is available upon request; the side-view videos have a resolution of 1920×1080 at either 60fps or 30fps; the top-down RGB videos have a resolution of 1280×720 at 15fps; the depth data has a resolution of 480×848, 32-bit floating-point format, with units in meters. This dataset is applicable to tasks such as robotic manipulation, imitation learning, action understanding, and depth estimation, with particular focus on challenges including tool use, bimanual coordination, and fine food manipulation. The full dataset is planned to expand to include 7 additional cooking tasks such as dumpling wrapping, dough kneading, deep frying, pan-searing, stewing, thickening with starch slurry, and marinating/seasoning, and task-level or action-level annotations can be provided. Full informed consent was obtained from all participants prior to data collection.
数据集概述:Chinese Commercial Kitchen Manipulation Dataset — Sample Pack v0.1
该数据集是亚洲首个真实商业厨房操作数据集,包含在中国一家正在营业的中餐馆内录制的专业厨师烹饪演示。
核心特点
- 场景与人员:采集于广东中山的真实商业中餐厅,由一位拥有约20年经验的专业厨师操作。
- 任务类型:涵盖真实的中餐烹饪任务,涉及工具使用、双手协调和精细食物操作。
- 视角与传感器:提供第一人称(自拍视角)、侧方和俯视多视角RGB视频,以及俯视视角的Intel RealSense D435I深度数据。
- 数据用途:专为机器人技术、具身智能、模仿学习和视觉动作理解等研究领域设计。
样本包内容
样本包包含两个示例任务的数据:
| 任务 | 中文名 | 包含视角 | 深度数据 |
|---|---|---|---|
| 任务1:切菜 | 切菜 | 3个视角(自拍、侧方、俯视) | ✅ 包含(HDF5格式,480×848,2379帧) |
| 任务2:翻炒 | 翻炒 | 2个视角(自拍、侧方) | ❌ 不包含(可申请获得) |
各视角视频规格:
- 自拍视角:头戴运动相机,预览版为1080p 30fps,原始采集为4K 30fps。
- 侧方视角:固定手机,任务1为1080p 60fps,任务2为1080p 30fps。
- 俯视RGB:RealSense D435I摄像头,1280×720,15fps。
- 俯视深度:RealSense D435I摄像头,848×480,HDF5文件,帧数2379,深度值单位为米(float32),无效像素值为65.535米,典型有效像素率约86%。
文件结构
样本包的文件目录结构如下:
task_01_cutting/ ├── egocentric_1080p.mp4 ├── side_view.mp4 ├── overhead/ │ ├── overhead.mp4 │ └── overhead_.PNG ├── depth/ │ ├── depth.hdf5 │ └── check_.jpg ├── egocentric_screenshot_.PNG └── side_view_.PNG task_02_stir_fry/ ├── egocentric_1080p.mp4 ├── side_view.mp4 ├── egocentric_screenshot_.PNG └── side_view_.PNG
- 深度数据读取(Python示例): python import h5py with h5py.File("task_01_cutting/depth/depth.hdf5", "r") as f: depth = f["depth_meters"][:] # (2379, 480, 848) float32, meters ts = f["timestamps"][:]
采集环境与许可
- 地点:中国广东省中山市一家正在营业的商业中餐馆。
- 参与者知情同意:所有参与者均签署了完整知情同意书。
- 许可证:该数据集采用 CC BY-NC 4.0 许可协议(非商业使用)。
完整数据与额外资源
公开样本包主要用于技术评估和早期研究反馈。可通过联系邮箱 andy@dynamicnova.com 申请获取完整数据,包括:
- 全4K自拍视频
- 更长的多视角MP4视频
- 额外的HDF5深度序列
- RealSense原始.bag文件
- 任务级或动作级标注
- 格式转换支持(HDF5、RLDS或LeRobot)
- 针对特定厨房流程的定制采集
完整数据集计划包括的任务有:切菜、翻炒、包饺子、揉面、油炸、煎、炖煮、勾芡、腌制/调味。





