DaivdYuan/hub-tennis-ball-basket-toss-lerobot

--- license: other tags: - robotics - lerobot - manipulation - unknown - nmotion-initiative --- # Tennis Ball Basket Toss ## Visualizers - [LeRobot Visualizer](https://huggingface.co/spaces/lerobot/visualize_dataset?path=%2FDaivdYuan%2Fhub-tennis-ball-basket-toss-lerobot) - [Neural Motion Visualizer](https://platform.neural-motion.org/open?repo=DaivdYuan%2Fhub-tennis-ball-basket-toss-lerobot) ## Overview This dataset is a conversion from an upstream robotics dataset into **LeRobot v3-compatible format**. It is intended to provide reproducible access in a unified schema. ## Source Dataset - Dataset ID: `hub-tennis-ball-basket-toss` - Project: `unknown` - Task: `Tennis Ball Basket Toss` - Upstream download/source URL: https://real.stanford.edu/umi-on-legs/tossing.zarr.zip - Upstream project page: https://umi-data.github.io/ - Contact: [e-mail] - Original paper: Fast-UMI: A Scalable and Hardware-Independent Universal Manipulation Interface (2024) - Notes: 500 demos; 1 envs; obs=Image,Proprio; actions=6 DoF End Effector,Parallel Gripper ## Dataset Format - Storage format: LeRobot v3 parquet-based dataset - Total episodes: 165 - Total frames: 194179 - FPS: 30 - Robot type: generic_robot - Observation feature: `observation.state` with shape `[7]` - Action feature: `action` with shape `[7]` - Video streams: present ## Video Modalities - `observation.images.camera0_rgb`: shape `[224, 224, 3]` - Frame rate: `30` FPS ## Conversion Notes - Target schema: LeRobot v3 - Conversion path: ingest -> preprocess -> zarr probe -> convert 1 episode -> upload - `action` fallback may be used when upstream `data/action` is missing. ## Dataset Structure `meta/info.json`: ```json { "codebase_version": "v3.0", "robot_type": "generic_robot", "total_episodes": 165, "total_frames": 194179, "fps": 30, "splits": {"train": "0:165"}, "data_path": "data/chunk-{chunk_index:03d}/file-{file_index:03d}.parquet", "video_path": "videos/{video_key}/chunk-{chunk_index:03d}/file-{file_index:03d}.mp4", "features": { "observation.state": {"dtype": "float32", "shape": [7]}, "action": {"dtype": "float32", "shape": [7]} } } ``` ## Data Splits - `{'train': '0:165'}` ## Citation This dataset was created using [LeRobot](https://github.com/huggingface/lerobot). Please also cite the original upstream project and paper listed above. ## Attribution - Original dataset credit remains with upstream creators. - Conversion and curation contribution: - **Convertor: David Yuan, [e-mail], Stanford University** - **NMotion Initiative** (opensource community): *we make embodied data accessible*--- license: 其他 tags: - 机器人学（robotics） - LeRobot（lerobot） - 操作任务（manipulation） - 未知（unknown） - NMotion Initiative（nmotion-initiative） --- # 网球投筐任务 ## 可视化工具 - [LeRobot可视化工具](https://huggingface.co/spaces/lerobot/visualize_dataset?path=%2FDaivdYuan%2Fhub-tennis-ball-basket-toss-lerobot) - [Neural Motion可视化工具](https://platform.neural-motion.org/open?repo=DaivdYuan%2Fhub-tennis-ball-basket-toss-lerobot) ## 概述 本数据集由上游机器人学数据集转换而来，适配**LeRobot v3格式**，旨在通过统一数据模式提供可复现的数据集访问途径。 ## 源数据集 - 数据集ID：`hub-tennis-ball-basket-toss` - 所属项目：未知 - 任务：网球投筐任务 - 上游下载/源地址：https://real.stanford.edu/umi-on-legs/tossing.zarr.zip - 上游项目页面：https://umi-data.github.io/ - 联系方式：[e-mail] - 原论文：《Fast-UMI：一种可扩展、硬件无关的通用操作接口》（2024） - 备注：共包含500段演示轨迹；1个实验环境；观测数据包含图像与本体感知信息；动作空间对应6自由度末端执行器与平行夹爪 ## 数据集格式 - 存储格式：基于Parquet的LeRobot v3数据集 - 总片段数：165 - 总帧数：194179 - 帧率：30 FPS - 机器人类型：通用机器人（generic_robot） - 观测特征：`observation.state`，形状为`[7]` - 动作特征：`action`，形状为`[7]` - 包含视频流 ## 视频模态 - 视频模态：`observation.images.camera0_rgb`，形状为`[224, 224, 3]` - 帧率：30 FPS ## 转换说明 - 转换目标数据模式：LeRobot v3 - 转换流程：数据摄入 → 预处理 → Zarr格式探测 → 单片段转换 → 上传 - 当上游`data/action`字段缺失时，可使用`action`字段作为后备 ## 数据集结构 `meta/info.json`： json { "codebase_version": "v3.0", "robot_type": "generic_robot", "total_episodes": 165, "total_frames": 194179, "fps": 30, "splits": {"train": "0:165"}, "data_path": "data/chunk-{chunk_index:03d}/file-{file_index:03d}.parquet", "video_path": "videos/{video_key}/chunk-{chunk_index:03d}/file-{file_index:03d}.mp4", "features": { "observation.state": {"dtype": "float32", "shape": [7]}, "action": {"dtype": "float32", "shape": [7]} } } ## 数据划分 数据划分规则：`{"train": "0:165"}`，即训练集包含索引0至165的所有片段 ## 引用信息 本数据集基于[LeRobot](https://github.com/huggingface/lerobot)构建。请同时引用上文提及的上游项目与原论文。 ## 署名声明 - 原始数据集的知识产权归上游创作者所有 - 转换与整理工作贡献： - **转换者：David Yuan，邮箱：[e-mail]，斯坦福大学** - **NMotion Initiative**（开源社区）：*我们致力于让具身智能数据触手可及*