video video 21 255 |
|---|
YAML Metadata Warning:The task_categories "robot-learning" is not in the official list: text-classification, token-classification, table-question-answering, question-answering, zero-shot-classification, translation, summarization, feature-extraction, text-generation, fill-mask, sentence-similarity, text-to-speech, text-to-audio, automatic-speech-recognition, audio-to-audio, audio-classification, audio-text-to-text, voice-activity-detection, depth-estimation, image-classification, object-detection, image-segmentation, text-to-image, image-to-text, image-to-image, image-to-video, unconditional-image-generation, video-classification, reinforcement-learning, robotics, tabular-classification, tabular-regression, tabular-to-text, table-to-text, multiple-choice, text-ranking, text-retrieval, time-series-forecasting, text-to-video, image-text-to-text, image-text-to-image, image-text-to-video, visual-question-answering, document-question-answering, zero-shot-image-classification, graph-ml, mask-generation, zero-shot-object-detection, text-to-3d, image-to-3d, image-feature-extraction, video-text-to-text, keypoint-detection, visual-document-retrieval, any-to-any, video-to-video, other
YAML Metadata Warning:The task_categories "imitation-learning" is not in the official list: text-classification, token-classification, table-question-answering, question-answering, zero-shot-classification, translation, summarization, feature-extraction, text-generation, fill-mask, sentence-similarity, text-to-speech, text-to-audio, automatic-speech-recognition, audio-to-audio, audio-classification, audio-text-to-text, voice-activity-detection, depth-estimation, image-classification, object-detection, image-segmentation, text-to-image, image-to-text, image-to-image, image-to-video, unconditional-image-generation, video-classification, reinforcement-learning, robotics, tabular-classification, tabular-regression, tabular-to-text, table-to-text, multiple-choice, text-ranking, text-retrieval, time-series-forecasting, text-to-video, image-text-to-text, image-text-to-image, image-text-to-video, visual-question-answering, document-question-answering, zero-shot-image-classification, graph-ml, mask-generation, zero-shot-object-detection, text-to-3d, image-to-3d, image-feature-extraction, video-text-to-text, keypoint-detection, visual-document-retrieval, any-to-any, video-to-video, other
Dataset Card for Grasp_OrangeToPlate_Dataset
本数据集包含 50 条遥操作演示记录,任务为使用 SO-101 机械臂将橘子抓取并放置到盘子中。数据采用 LeRobot v3.0 格式存储,包含关节位置、动作指令和前视摄像头视频。
Dataset Details
Dataset Description
| 属性 | 值 |
|---|---|
| 格式 | LeRobot v3.0 |
| 帧率 | 30 FPS |
| 总集数 | 50 集 (episodes) |
| 总帧数 | 28,132 帧 |
| 总时长 | 约 15.6 分钟 |
| 机器人 | SO-100 Follower 机械臂 |
| 任务数 | 1 |
| 语言 | 中文 |
| 许可 | MIT |
关节定义(action 和 observation.state 均为 6 维,单位:度):
| 索引 | 关节名称 | 范围 |
|---|---|---|
| 0 | shoulder_pan.pos |
[-58°, 28°] |
| 1 | shoulder_lift.pos |
[-110°, 58°] |
| 2 | elbow_flex.pos |
[-94°, 97°] |
| 3 | wrist_flex.pos |
[7°, 103°] |
| 4 | wrist_roll.pos |
[-21°, 37°] |
| 5 | gripper.pos |
[0°, 72°] |
摄像头:
| 名称 | 分辨率 | 编码 | 说明 |
|---|---|---|---|
observation.images.front |
1920x1080 | AV1 (yuv420p, CRF 30) | 前视摄像头,Full HD |
Dataset Sources
- 仓库: [More Information Needed]
- 论文 [可选]: [More Information Needed]
- 演示 [可选]: [More Information Needed]
Uses
Direct Use
本数据集适用于以下场景:
- 模仿学习训练:使用 ACT、Diffusion、VQ-BeT 等策略训练机械臂抓取放置任务
- 策略评估:在 SO-100 机械臂上测试训练好的策略
- 数据格式参考:作为 LeRobot v3.0 格式数据集的参考示例
- 遥操作数据分析:分析人类遥操作的动作特征和跟踪误差
Out-of-Scope Use
- 本数据集仅包含单一任务(抓取橘子放到盘子),不适用于泛化到其他物体或任务
- 数据来自 SO-100 机械臂,关节范围和运动学特性与其他机器人不同,直接迁移需谨慎
- 数据集仅包含前视摄像头,缺少深度或多视角信息,不适用于需要 3D 感知的方法
Dataset Structure
目录结构
Grasp_OrangeToPlate_Dataset/
├── data/
│ └── chunk-000/
│ └── file-000.parquet # 主数据文件 (28,132 行, 7 列)
├── meta/
│ ├── info.json # 数据集全局信息 (schema, fps, 机器人类型等)
│ ├── stats.json # 全局统计 (min, max, mean, std, 分位数)
│ ├── tasks.parquet # 任务定义 (1 个任务)
│ └── episodes/
│ └── chunk-000/
│ └── file-000.parquet # 每集元数据 (50 行, 93 列)
└── videos/
└── observation.images.front/
└── chunk-000/
├── file-000.mp4 # 第 0-8 集
├── file-001.mp4 # 第 9-21 集
├── file-002.mp4 # 第 22-36 集
├── file-003.mp4 # 第 37-48 集
└── file-004.mp4 # 第 49 集
数据字段
**主数据 (data/chunk-000/file-000.parquet)**:
| 字段 | 类型 | 说明 |
|---|---|---|
action |
float32[6] | 目标关节位置(遥操作主臂的 commanded 位置) |
observation.state |
float32[6] | 实际关节位置(从臂的 measured 位置) |
timestamp |
float32[1] | 时间戳(秒,从每集 0 开始) |
frame_index |
int64[1] | 帧索引(每集内从 0 开始) |
episode_index |
int64[1] | 集索引(0-49) |
index |
int64[1] | 全局帧索引 |
task_index |
int64[1] | 任务索引(始终为 0) |
**视频 (videos/observation.images.front/)**:
- 格式:MP4 (AV1 编码, yuv420p, 30 FPS)
- 分辨率:1920x1080
- CRF: 30, GOP: 2, preset: 12
- 像素值范围:归一化到 [0, 1],各通道均值约 0.34-0.35
数据划分:
| 划分 | 集范围 | 数量 |
|---|---|---|
train |
0-49 | 50 集 |
每集长度分布
| 统计量 | 值 |
|---|---|
| 最小长度 | 2 帧(第 17 集,异常数据) |
| 最大长度 | 900 帧(约 30 秒) |
| 平均长度 | 562.6 帧(约 18.8 秒) |
| 中位数长度 | 523.5 帧(约 17.5 秒) |
注意:第 17 集仅包含 2 帧(0.067 秒),为录制失败的异常数据,训练时建议排除。
Dataset Creation
Curation Rationale
本数据集用于记录和训练 SO-100 机械臂完成"把橘子放到盘子里面"的抓取放置任务。通过遥操作方式采集人类演示数据,用于后续模仿学习策略的训练和评估。
Source Data
数据采集与处理
- 采集方式:遥操作(Teleoperation),使用 SO-100 Leader-Follower 主从臂配置
- 帧率:30 FPS
- 数据格式:LeRobot v3.0(Parquet + 视频分块存储)
- 关节单位:度(degrees)
- 视频编码:AV1,CRF 30,用于高效存储
action 与 observation.state 的区别:action 是主臂(Leader)的目标关节位置指令,observation.state 是从臂(Follower)的实际测量关节位置。两者之间存在跟踪误差,最大可达约 5 度,这是 SO-100 主从臂系统的典型特性。
数据生产者
数据由操作 SO-100 Leader-Follower 机械臂的操作员通过遥操作方式采集。
Personal and Sensitive Information
本数据集不包含任何个人或敏感信息。数据仅包含机械臂关节位置和摄像头拍摄的物体(橘子、盘子)画面。
Bias, Risks, and Limitations
已知限制
- 单一任务:仅包含"把橘子放到盘子里面"一个任务,无法泛化到其他物体或动作
- 单一场景:所有数据在固定环境下采集,光照、桌面布局、橘子大小等条件变化有限
- 机器人特定:数据针对 SO-100 机械臂的关节范围和运动学特性,迁移到其他机器人需要适配
- 数据量较小:50 集演示对于复杂策略训练可能不足,建议与其他数据集配合使用或进行数据增强
- 异常数据:第 17 集仅 2 帧,为录制失败的异常数据
Recommendations
建议使用者注意以下事项:
- 训练前排除第 17 集(仅 2 帧的异常数据)
- 注意
action(指令位置)和observation.state(实际位置)之间的跟踪误差 - 在目标环境中充分评估策略性能后再部署到真实机器人
- 本数据集适合用作入门示例或小规模验证,生产环境建议采集更多样化的数据
Citation
BibTeX:
[More Information Needed]
APA:
[More Information Needed]
More Information
- LeRobot 文档: https://github.com/huggingface/lerobot
- LeRobot v3.0 数据格式规范: https://github.com/huggingface/lerobot
Dataset Card Contact
[More Information Needed]
- Downloads last month
- 48