yky666/humanoid_long_horizon_tasks

Humanoid Long-Horizon Tasks

人形机器人长时序操作任务演示数据集。仓库按任务提供独立的数据归档，供机器人操作、模仿学习与具身智能研究使用。

This dataset provides task-specific demonstration archives for humanoid robot manipulation research. Each task is distributed as a separate archive.

数据包 / Task packages

任务 / Task	说明 / Description	文件 / File	格式 / Format	大小 / Size
food_sorting	食品分类整理 / Food sorting	food_sorting.tar	TAR	约 54.04 GB
pick_water_bottle_handover_and_insert	水瓶抓取、递交与插入 / Pick up, hand over, and insert a water bottle	water_bottle_pick_and_insert.rar	RAR	约 2.68 GB

仓库结构 / Repository layout

.
├── .gitattributes
├── README.md
├── food_sorting.tar
└── water_bottle_pick_and_insert.rar

每个归档对应一个任务数据包。当前仓库不包含解压后的 episode 文件，也没有声明适用于所有任务的统一字段结构或 Hugging Face datasets 加载器；请根据所选任务包内的实际文件读取数据。

Each archive is a separate task package. The repository does not include extracted episode files or define a shared schema/loader across tasks. Inspect the selected archive before parsing it.

下载 / Download

使用 Hugging Face CLI 下载单个任务包（只运行需要的命令）：

hf download yky666/humanoid_long_horizon_tasks food_sorting.tar --repo-type dataset --local-dir data

hf download yky666/humanoid_long_horizon_tasks water_bottle_pick_and_insert.rar --repo-type dataset --local-dir data

也可以在 Files and versions 页面选择对应文件下载。food_sorting.tar 约 54 GB；下载和解压前请为归档及解压结果预留足够磁盘空间。

Use the Hugging Face CLI to download only the task archive you need. The food-sorting archive is about 54 GB; allow additional disk space for extraction.

解压 / Extract

TAR 文件可用系统 tar 命令解压（以下命令以 Bash、macOS/Linux 或 WSL 为例）：

mkdir -p data/food_sorting
tar -xf data/food_sorting.tar -C data/food_sorting

RAR 文件需要安装 7-Zip（或兼容的解压工具）。以下命令在 Bash、macOS/Linux 或 WSL 中运行：

mkdir -p data/water_bottle_pick_and_insert
7z x data/water_bottle_pick_and_insert.rar -odata/water_bottle_pick_and_insert

PowerShell 用户可用以下命令解压 TAR：

New-Item -ItemType Directory -Force .\data\food_sorting | Out-Null
tar -xf .\data\food_sorting.tar -C .\data\food_sorting

RAR 可在 PowerShell 中通过 7-Zip 解压：

7z x .\data\water_bottle_pick_and_insert.rar -o.\data\water_bottle_pick_and_insert

每个归档解压到独立目录。以上命令只在本地创建目录，不会改变 Hugging Face 仓库中的文件布局。

Extract each archive into its own directory. These commands create local folders; they do not change the layout on the Hub.

许可 / License

本数据集按 Apache License 2.0 发布。使用或再分发时请遵循许可证条款，并注明数据集来源：https://huggingface.co/datasets/yky666/humanoid_long_horizon_tasks。

This dataset is released under the Apache License 2.0. Follow the license terms and cite the dataset source when using or redistributing it.

G1D food sorting: subtask annotations and reproducible split

Raw food_sorting.tar is preserved. Additional annotations and processed LeRobot datasets are under g1d_food_sorting/. Videos in archives are dereferenced (no local symlink dependencies).

• annotations/: 197 episode labels, seven task prompts, grouped split manifest, and fixed validation frames.
• processed/whole_197.tar: complete episodes.
• processed/subtasks_all.tar: all five-stage segments using seven prompts.
• processed/subtasks_train.tar: 177 source episodes / 885 segments / 183647 frames.
• processed/subtasks_validation.tar: 20 source episodes / 100 segments / 21907 frames.

Split by source episode, never by independent segment. Three RGB views; 30Hz; 16D absolute joint states/actions (left arm7, right arm7, left gripper, right gripper). Lift/waist are external controls. Normalization is fitted to training data only.

Code, pinned downloads and experiment details: g1d_experiments (private repository; permission required). Large archives may be uploading during initial publication; completion is recorded in its ARTIFACTS.json.