Organize bilingual dataset into real-world and simulation subsets

Rename the real-world full-shot training directory, reserve the other three subset directories, document joint and end-effector action fields, and remove operational notes from current dataset documentation.

README.md

@@ -1,73 +1,87 @@
 ---
-pretty_name: Value Model Benchmark Training Data (Real-World Full-Shot)
+pretty_name: Value Model Benchmark Training Data
 language:
 - en
 - zh
 tags:
 - robotics
 - lerobot
-- value-model-benchmark-training-data
+- value-model-benchmark
 - training-data
 - real-world
 - full-shot
-- real-world-full-shot-data
 ---
 
 # Value Model Benchmark Training Data
 
-**Current release / 当前发布：`real-world-full-shot-data`**
+[中文](#中文) · [English](#english)
 
-[中文说明](#中文说明) · [English](#english) · [Download and loading / 下载与读取](#download-and-loading)
+## 中文
 
-## 中文说明
+本数据集面向机器人 Value Model Benchmark 训练，按**真实世界（Real-World）**和**仿真世界（Simulation）**两个大类组织；每个大类再分为 **Full-Shot** 与 **One-Shot** 两种数据设置。
 
-本仓库用于发布 Value Model Benchmark 的机器人轨迹训练数据，按**采集域（真机 / 仿真）**与**数据使用设置（full-shot / one-shot）**区分子集。
+当前训练数据为 **Real-World Full-Shot Training Data（真机全样本训练集）**，子集标识为 **`real-world-full-shot-data`**：包含 G2A 机器人的 **20 个任务、2,000 条轨迹，每个任务 100 条**。其他三个子类保留独立扩展位置，不计入这 2,000 条轨迹。
 
-**当前只有真机 full-shot 数据：G2A 机器人、20 个任务、2,000 条完整轨迹，每个任务 100 条。当前不包含 one-shot 或仿真数据。**“2,000 条”指 2,000 个 episode（轨迹），不是 2,000 个任务或 2,000 帧。
+### 1. 真实世界 / Real-World
 
-### 子集命名与发布范围 / Subset catalog
-
-| 子集标识 / Subset ID | 采集域 / Domain | 设置 / Setting | 当前状态 / Availability |
+| 子类 | 子集标识 | 数据内容 | 数据目录 |
 | --- | --- | --- | --- |
-| `real-world-full-shot-data` | 真机 / Real-world | Full-shot | **已上传 / Uploaded：20 tasks × 100 episodes = 2,000 episodes** |
-| `real-world-one-shot-data` | 真机 / Real-world | One-shot | 计划发布，当前不可用 / Planned, not available |
-| `simulation-full-shot-data` | 仿真 / Simulation | Full-shot | 计划发布，当前不可用 / Planned, not available |
-| `simulation-one-shot-data` | 仿真 / Simulation | One-shot | 计划发布，当前不可用 / Planned, not available |
+| **Full-Shot** | `real-world-full-shot-data` | **G2A；20 个任务；2,000 条轨迹** | [`data/real-world/full-shot/`](./data/real-world/full-shot/) |
+| One-Shot | `real-world-one-shot-data` | 扩展位置 | [`data/real-world/one-shot/`](./data/real-world/one-shot/) |
+
+### 2. 仿真世界 / Simulation
 
-本次 **full-shot** 标识表示提供当前已发布的每个任务全部 100 条轨迹。后续 one-shot 子集将单独说明轨迹选择方法、来源映射及实验使用规则。full-shot 不表示原始采集资产已全部公开，也不表示所有轨迹都有已核验的成功标签。
+| 子类 | 子集标识 | 数据内容 | 数据目录 |
+| --- | --- | --- | --- |
+| Full-Shot | `simulation-full-shot-data` | 扩展位置 | [`data/simulation/full-shot/`](./data/simulation/full-shot/) |
+| One-Shot | `simulation-one-shot-data` | 扩展位置 | [`data/simulation/one-shot/`](./data/simulation/one-shot/) |
 
-`real-world-full-shot-data` 是当前子集的明确标识。为保持已有下载路径、校验和及已完成的全量验收有效，**本次只更新仓库级文档和标识，不移动或修改训练文件**。实际数据路径仍为 `data/ValueModelBenchmark_lerobot_v30_video/`。[`subset_catalog.json`](./subset_catalog.json) 记录子集标识、实际路径和数据版本；后续子集使用独立目录，不覆盖现有真机 full-shot 数据。
+Full-shot 表示使用当前子集每个任务的完整轨迹集合；本真机训练集为每个任务 100 条。One-shot 子集的样本选择方法与来源对应关系由各自的数据说明定义。数据分类、规模与路径也记录在 [`subset_catalog.json`](./subset_catalog.json) 中。
 
-### 当前 2,000 条轨迹概况 / Release at a glance
+### 真机 Full-Shot 训练集概况
 
-| 项目 / Item | 当前发布内容 / Current release |
+| 项目 | 内容 |
 | --- | --- |
-| 数据类型 / Domain | 真实机器人操作轨迹 / Real-world robot manipulation |
-| 机器人 / Robot | G2A |
-| 任务 / Tasks | 20 |
-| 轨迹 / Episodes | 2,000；每个任务 100 条 / 100 per task |
-| 总帧数 / Frames | 1,530,987 个轨迹时间步 / trajectory time steps |
-| 采样率 / Frame rate | 30 FPS |
-| 总轨迹时长 / Total duration | 约 / approximately 14.18 hours |
-| 单条轨迹时长 / Episode duration | 平均 / mean 25.52 s；范围 / range 11.37–66.03 s |
-| 相机 / Cameras | 头部、左腕、右腕 3 路 RGB / head, left wrist, right wrist；6,000 MP4 files |
-| 头部图像 / Head image | 640 × 400 px；CHW `[3, 400, 640]` |
-| 左右腕图像 / Wrist images | 各 / each 640 × 528 px；CHW `[3, 528, 640]` |
-| 状态 / State | `float32[290]` |
-| 动作 / Action | `float32[52]` |
-| 格式 / Format | LeRobot v3.0：Parquet + HEVC video + metadata / provenance |
-| 冻结数据包 / Frozen bundle | 8,010 files；23,226,080,803 bytes，约 / approximately 23.23 GB |
-| 当前 split / Published split | `train: 0:2000`；没有独立 validation / test split / none published |
-
-时长以 episode `length / 30` 汇总；帧数与时长不按 3 路相机重复计数。数据包规模不包含仓库级文档。任务覆盖物体放置与分类、按钮与键盘操作、桌面和家居整理、抽屉操作、厨房与餐饮操作、包装及装饰摆放。
-
-Durations are computed from episode `length / 30`; neither frames nor durations are multiplied by the three camera views. Bundle size excludes repository-level documentation. Tasks cover object placement and sorting, buttons and keypads, desk and household organization, drawers, kitchen activities, packaging, and decorative placement.
-
-### 任务清单 / Task inventory
-
-英文任务名与 `meta/tasks.parquet` 一致；中文为辅助翻译，源任务名中的数字后缀保留。English instructions match `meta/tasks.parquet`; Chinese descriptions are reading aids, and original numeric suffixes are retained.
-
-| `task_index` | Source task ID | English task instruction | 中文任务说明 | Episodes | Frames |
+| 采集域 / 机器人 | 真实世界 / G2A |
+| 任务与轨迹 | 20 个任务；2,000 条轨迹；每个任务 100 条 |
+| 总帧数 | 1,530,987 个轨迹时间步 |
+| 采样率 | 30 FPS |
+| 轨迹时长 | 总计约 14.18 小时；平均 25.52 秒；范围 11.37–66.03 秒 |
+| 视觉 | 头部、左腕、右腕三路 RGB 视频，共 6,000 个 MP4 文件 |
+| 图像分辨率 | 头部 640 × 400；左 / 右腕各 640 × 528；元数据使用 CHW |
+| 状态 | `observation.state`：`float32[290]` |
+| 动作 | `action`：`float32[52]`，**同时包含关节目标与末端目标位姿** |
+| 格式 | LeRobot v3.0；Parquet 数值数据 + HEVC 视频 + 元数据 |
+| 压缩数据规模 | 约 23.23 GB |
+| 数据划分 | `train: 0:2000` |
+
+一条数据指一个完整 episode（轨迹），不是一个任务或一帧。轨迹时长按 `length / 30` 计算；帧数与时长不按三路相机重复计数。任务覆盖物体放置与分类、按钮与键盘操作、桌面与家居整理、抽屉操作、厨房操作、包装与装饰摆放。
+
+### 动作到底是关节还是末端位姿？
+
+**两者都有。**LeRobot 是数据组织格式，不要求动作一定是关节空间或末端空间。本数据的 52 维 `action` 是一个组合向量，具体由源数据字段定义，而不是 52 个机器人关节。
+
+| `action` 切片（Python，右端不含） | 维度 | 含义 |
+| --- | ---: | --- |
+| `[0:2]` | 2 | 左 / 右执行器目标开度或行程；保留源数值 |
+| `[2:8]` | 6 | 双臂末端目标位置；每臂 XYZ，单位 m，坐标系 `base_link` |
+| `[8:16]` | 8 | 双臂末端目标朝向；每臂四元数 XYZW，坐标系 `base_link` |
+| `[16:30]` | 14 | 双臂关节目标角度，单位 rad；关节顺序见 `meta/info.json` |
+| `[30:33]` | 3 | 头部关节目标角度，单位 rad |
+| `[33:38]` | 5 | 腰部关节目标角度，单位 rad |
+| `[38:44]` | 6 | 底盘目标速度；XYZ 线速度 + XYZ 角速度 |
+| `[44:48]` | 4 | 左手参考姿态四元数 XYZW；不同于上述末端目标朝向 |
+| `[48:52]` | 4 | 右手参考姿态四元数 XYZW；不同于上述末端目标朝向 |
+
+**末端目标位置与朝向由关节 action 通过正运动学（FK）计算，不是另一套独立的遥操作末端指令。**元数据里部分字段声明 `dimensions: 0` 且 `indices: []`，这些字段不占用 52 维向量，也不代表存在额外动作数据。
+
+290 维 `observation.state` 同样包含关节与末端信息，例如双臂关节位置 `[108:122]`、速度 `[136:150]`、力矩 `[122:136]`，以及 `base_link` 下的双臂末端位置 `[26:32]`、朝向 `[56:64]`。完整字段、单位、坐标系和索引见子集内的 `meta/info.json`。
+
+### 任务清单 / Task Inventory
+
+英文任务名与 `meta/tasks.parquet` 一致；中文为辅助翻译，保留源任务名的数字后缀。
+
+| `task_index` | Source task ID | English instruction | 中文说明 | Episodes | Frames |
 | --- | --- | --- | --- | ---: | ---: |
 | 0 | 16249 | Put colored blocks in the bowl and pour | 将彩色积木放入碗中并倒出 | 100 | 57,018 |
 | 1 | 16360 | Press the numeric keypad | 按数字键盘 | 100 | 59,414 |
@@ -89,139 +103,105 @@ Durations are computed from episode `length / 30`; neither frames nor durations
 | 17 | 31119 | Logistics packaging | 物流包装 | 100 | 71,543 |
 | 18 | 31913 | Floral arrangement and flameless aromatherapy placement | 插花与无火香薰摆放 | 100 | 98,090 |
 | 19 | 37438 | Prepare breakfast | 准备早餐 | 100 | 88,685 |
-| **Total / 合计** | — | **20 tasks / 20 个任务** | — | **2,000** | **1,530,987** |
+| **合计 / Total** | — | **20 tasks** | — | **2,000** | **1,530,987** |
 
-### 内容、使用边界与发布状态
+### 数据使用说��
 
-- 每条轨迹包含任务关联、逐帧状态 / 动作、时间戳及 3 路相机视频。`meta/episodes/` 定义轨迹长度、文件位置和视频区间；`provenance/episode_mapping.json` 提供源任务和轨迹映射。
-- 保留原始数值、单位、坐标系、字段索引和 XYZW 四元数顺序，仅重建全局任务 / 轨迹 / 帧索引。夹爪值不归一化；辅助字段 `ts`、`action_src_status` 保留，不推断未核实的语义。
-- 原始 HEVC 视频逐字节复制，不重编码、缩放或改变帧率。每条轨迹对应 1 个数值 Parquet 文件和 3 个 MP4 文件。数值统计和全局分位数精确计算；图像统计使用 provenance 中声明的时间 / 空间采样，不包含图像分位数。
-- **当前未声称已提供或核验奖励、成功 / 失败、进度或 value target 标签。**full-shot 是子集设置标识，不是新增逐轨迹语义标注。兼容的是 LeRobot 文件组织，不是 RoboDojo 的 14 维动作 / 状态空间，跨机器人使用需另行适配。
-- 截至 **2026-10-14**，2,000 条真机轨迹已上传，仓库公开且可匿名访问；100 条 pilot 远端读取验收已通过，**最终全量远端验收已通过**：全部数值行、轨迹边界、6,000 个视频逐字节一致性及每条轨迹首中末帧检查通过。
-- 数据许可尚未指定；公开可见不等于已明确使用授权，不继承 RoboDojo 许可。HDF5 尚未发布；从本 LeRobot 数据重建的 HDF5 将明确标为派生归档，不能称为原始传感器录制。
-- 未经审查的辅助侧文件、用户标识、机器人序列号、内部地址、凭证和日志不在发布内容中。冻结导出文档 / provenance 保留导出时“公开待授权”的审批快照，当前可见性以本仓库级说明为准；本次不重写历史记录和校验和。
+原始数值、单位、坐标系与四元数顺序保留；夹爪数值未归一化。HEVC 视频保持原始编码、分辨率和帧率。每条轨迹包含一个数值 Parquet 文件及三路视频；轨迹边界由 `meta/episodes/` 定义，来源对应关系见 `provenance/episode_mapping.json`。
+
+Full-shot 是样本设置，不是成功 / 失败标注。数据集用途不代表额外提供了 reward、progress 或 value target 标签。LeRobot 格式兼容不等于机器人动作空间兼容；使用模型时应按实际字段选择关节或末端表示。数据许可尚未指定，使用前请确认适用许可。
 
 ## English
 
-This repository publishes robot trajectory training data for the Value Model Benchmark, organized by **collection domain (real-world / simulation)** and **data-use setting (full-shot / one-shot)**.
+This dataset provides robot trajectory training data for the Value Model Benchmark. It is organized into **Real-World** and **Simulation** domains, each with **Full-Shot** and **One-Shot** subsets.
 
-**Only `real-world-full-shot-data` is currently available: 2,000 G2A real-robot episodes across 20 tasks, with 100 episodes per task. No one-shot or simulation data is included.** An episode is one trajectory, not one task or one frame. The bilingual tables above provide release statistics and all 20 task instructions.
+The current training set is **Real-World Full-Shot Training Data**, identified as **`real-world-full-shot-data`**: **2,000 G2A episodes across 20 tasks, with 100 episodes per task**. Separate extension directories are reserved for the other three subsets; they are not part of these 2,000 episodes.
 
-Here, **full-shot** identifies the release containing all 100 published episodes per task. Future one-shot data will document selection procedures, source mappings, and experimental usage separately. Full-shot does not claim that all original acquisition assets are public or that every episode has a verified success label.
+### 1. Real-World
 
-The subset ID is `real-world-full-shot-data`; the actual storage path remains **`data/ValueModelBenchmark_lerobot_v30_video/`**. This documentation-only relabeling does not move or modify training files, change bundle checksums, or invalidate existing paths and completed verification. [`subset_catalog.json`](./subset_catalog.json) maps subset IDs to actual paths and data revisions. Planned `real-world-one-shot-data`, `simulation-full-shot-data`, and `simulation-one-shot-data` are not yet available and will not overwrite this release.
+| Setting | Subset ID | Contents | Directory |
+| --- | --- | --- | --- |
+| **Full-Shot** | `real-world-full-shot-data` | **G2A; 20 tasks; 2,000 episodes** | [`data/real-world/full-shot/`](./data/real-world/full-shot/) |
+| One-Shot | `real-world-one-shot-data` | Reserved for extension | [`data/real-world/one-shot/`](./data/real-world/one-shot/) |
 
-The release follows LeRobot v3.0 file organization while preserving the source embodiment, dimensions, frame rate, image resolution, and action semantics. Each episode has one numeric Parquet file and three byte-for-byte copied HEVC videos. Original numerical values, units, coordinate frames, field-index mappings, and XYZW quaternion ordering are retained; gripper values are not normalized. Tabular statistics and global quantiles are exact; image statistics use documented temporal/spatial sampling without image quantiles.
+### 2. Simulation
 
-## Release status
+| Setting | Subset ID | Contents | Directory |
+| --- | --- | --- | --- |
+| Full-Shot | `simulation-full-shot-data` | Reserved for extension | [`data/simulation/full-shot/`](./data/simulation/full-shot/) |
+| One-Shot | `simulation-one-shot-data` | Reserved for extension | [`data/simulation/one-shot/`](./data/simulation/one-shot/) |
 
-- Repository and dataset documentation initialized on **2026-10-14**.
-- Training release: **LeRobot v3.0 video format**, **20 tasks, 2,000 episodes, 1,530,987 frames**, **8,010 files / 23,226,080,803 bytes** uploaded. Final full-release remote verification has **passed**.
-- Pilot scope: **1 task, 100 episodes, 57,018 frames**. Upload and remote loader verification have **passed**.
-- Verified full source scope: **20 tasks, 2,000 episodes, 1,530,987 frames**, approximately **23.30 GB** of existing compressed LeRobot files. This is not a claim about the completeness of the original sensor recordings.
-- HDF5 archival export: **planned; not yet uploaded**.
-- Visibility: **public**, authorized by the dataset owner on **2026-10-14**. Anonymous access is available. A data license has not yet been specified; no license is inferred from RoboDojo.
+Full-shot uses the complete episode collection within each task of the subset; this real-world set contains 100 episodes per task. Each one-shot release defines its selection procedure and source mapping in its own documentation. [`subset_catalog.json`](./subset_catalog.json) records subset identities, sizes, and paths.
 
-The repository title describes the intended use. It does **not** imply that rewards, success/failure labels, progress labels, or value targets have already been supplied or verified. The full-shot designation is a subset setting, not new per-episode semantic annotation.
+### Real-World Full-Shot Overview
 
-Frozen export documentation/provenance retains the approval snapshot from export time, which may describe public authorization as pending. Current visibility is described by this repository-level card; this relabeling does not rewrite historical records or frozen checksums. No data license has been specified; public visibility does not establish licensing terms.
+| Item | Specification |
+| --- | --- |
+| Domain / embodiment | Real-world robot manipulation / G2A |
+| Tasks / episodes | 20 tasks / 2,000 episodes / 100 episodes per task |
+| Frames / frame rate | 1,530,987 trajectory time steps / 30 FPS |
+| Duration | Approximately 14.18 hours total; mean 25.52 s; range 11.37–66.03 s |
+| Vision | Head, left wrist, right wrist RGB; 6,000 MP4 files |
+| Resolution | Head: 640 × 400; each wrist: 640 × 528; metadata uses CHW |
+| State | `observation.state`: `float32[290]` |
+| Action | `action`: `float32[52]`, containing **both joint targets and end-effector target poses** |
+| Format / size | LeRobot v3.0; Parquet + HEVC video + metadata; approximately 23.23 GB |
+| Split | `train: 0:2000` |
 
-## Repository layout
+An episode is one complete trajectory, not one task or one frame. Durations use `length / 30`; frame counts and durations are not multiplied by the three cameras. The bilingual task inventory above preserves the exact instructions in `meta/tasks.parquet`, including numeric suffixes.
 
-Only the available real-world full-shot subset is shown below. Planned one-shot, simulation, and HDF5 directories do not yet exist:
+### Joint Actions or End-Effector Poses?
 
-```text
-README.md
-subset_catalog.json
-data/
-└── ValueModelBenchmark_lerobot_v30_video/
-    ├── README.md
-    ├── data/chunk-*/file-*.parquet
-    ├── videos/
-    │   ├── observation.images.cam_high/chunk-*/file-*.mp4
-    │   ├── observation.images.cam_left_wrist/chunk-*/file-*.mp4
-    │   └── observation.images.cam_right_wrist/chunk-*/file-*.mp4
-    ├── meta/
-    │   ├── info.json
-    │   ├── stats.json
-    │   ├── tasks.parquet
-    │   └── episodes/chunk-*/file-*.parquet
-    ├── provenance/
-    └── checksums.sha256
-```
+**Both are present.**LeRobot specifies a data format, not a mandatory control space. The 52-dimensional action is a composite vector, not a vector of 52 joint angles.
 
-The directory above maps to `real-world-full-shot-data`; its existing name is retained for download and verification compatibility.
+| Python slice, stop-exclusive | Dimensions | Meaning |
+| --- | ---: | --- |
+| `[0:2]` | 2 | Left/right effector target opening or travel; original values retained |
+| `[2:8]` | 6 | Two arm end-effector target XYZ positions, in m, in `base_link` |
+| `[8:16]` | 8 | Two arm end-effector target orientations, XYZW quaternions, in `base_link` |
+| `[16:30]` | 14 | Two arm joint target positions, in rad; joint ordering is defined in `meta/info.json` |
+| `[30:33]` | 3 | Head joint target positions, in rad |
+| `[33:38]` | 5 | Waist joint target positions, in rad |
+| `[38:44]` | 6 | Base target velocity: XYZ linear + XYZ angular velocity |
+| `[44:48]` | 4 | Left reference-pose XYZW quaternion, distinct from end-effector target orientation |
+| `[48:52]` | 4 | Right reference-pose XYZW quaternion, distinct from end-effector target orientation |
 
-## LeRobot v3.0 training schema
+**End-effector target positions and orientations are computed by forward kinematics from joint actions; they are not an independent set of teleoperation end-effector commands.**Fields declared with `dimensions: 0` and `indices: []` do not occupy the action vector.
 
-| Feature | Source-preserving specification |
-| --- | --- |
-| `observation.state` | `float32[290]`; original feature descriptions and index mappings retained |
-| `action` | `float32[52]`; original feature descriptions and index mappings retained |
-| `timestamp` | `float32`; episode-relative frame timestamp |
-| `ts` | `int64[1]`; retained source auxiliary field |
-| `action_src_status` | `float32[1]`; retained source auxiliary field |
-| `frame_index` | `int64`; frame index within the episode |
-| `episode_index` | `int64`; release-local contiguous episode index |
-| `index` | `int64`; release-local contiguous global frame index |
-| `task_index` | `int64`; release-local task index linked to `meta/tasks.parquet` |
-| `observation.images.cam_high` | Video; metadata shape `[3,400,640]` |
-| `observation.images.cam_left_wrist` | Video; metadata shape `[3,528,640]` |
-| `observation.images.cam_right_wrist` | Video; metadata shape `[3,528,640]` |
-
-The source uses **30 FPS** and **HEVC** video. Source-only auxiliary columns, including `ts` and `action_src_status`, must not be silently dropped; their treatment is recorded in the conversion provenance.
-
-Camera names are mapped as follows:
+The 290-dimensional observation state also includes both representations: arm joint positions `[108:122]`, velocities `[136:150]`, and torques `[122:136]`, plus end-effector positions `[26:32]` and XYZW orientations `[56:64]` in `base_link`. Full field definitions, units, coordinate frames, and indices are in the subset's `meta/info.json`.
 
-```text
-observation.images.top_head   -> observation.images.cam_high
-observation.images.hand_left  -> observation.images.cam_left_wrist
-observation.images.hand_right -> observation.images.cam_right_wrist
-```
+### Usage Notes
 
-These aliases align the public feature names with RoboDojo; they do not change camera placement, calibration, image contents, or the meaning of a measurement. The full release copies all three MP4 streams without re-encoding.
+Original numerical values, units, coordinate frames, and quaternion order are retained; gripper values are not normalized. HEVC videos retain their original encoding, resolution, and frame rate. Each episode has one numeric Parquet file and three camera videos. Episode boundaries are defined in `meta/episodes/`; source mappings are in `provenance/episode_mapping.json`.
 
-`meta/info.json` declares `codebase_version: v3.0`, `robot_type: G2A`, `fps: 30`, feature dtypes/shapes, and the actual release counts. The full training split is `{"train": "0:2000"}`. Frame and episode indices are rebuilt consistently when merging tasks. `meta/episodes/` contains episode lengths, task associations, Parquet locations, and camera-specific video offsets. `meta/stats.json` is recomputed for the exported release.
+Full-shot is a sampling setting, not a success/failure annotation. Intended use does not imply additional reward, progress, or value-target labels. Format compatibility is not action-space compatibility: select the joint or end-effector fields appropriate to your model. No data license has been specified; confirm applicable terms before use.
 
-The path templates match RoboDojo:
+## Repository Layout / 仓库目录
 
 ```text
-data/chunk-{chunk_index:03d}/file-{file_index:03d}.parquet
-videos/{video_key}/chunk-{chunk_index:03d}/file-{file_index:03d}.mp4
+README.md
+subset_catalog.json
+data/
+├── real-world/
+│   ├── full-shot/
+│   │   ├── README.md
+│   │   ├── data/chunk-*/file-*.parquet
+│   │   ├── videos/<camera>/chunk-*/file-*.mp4
+│   │   ├── meta/
+│   │   ├── provenance/
+│   │   └── checksums.sha256
+│   └── one-shot/
+│       └── README.md
+└── simulation/
+    ├── full-shot/
+    │   └── README.md
+    └── one-shot/
+        └── README.md
 ```
 
-Shard occupancy is not itself a compatibility guarantee. The current full release uses one episode per numeric data file and per camera video to preserve the original media; episode segmentation is always defined by metadata.
+## Download and Loading / 下载与读取
 
-## Compatibility boundary
-
-**RoboDojo-compatible file organization does not mean identical robot tensors.** RoboDojo's reference joint-space release uses 14-dimensional state/action, 25 FPS, 640×480 images, and AV1 video. This source is G2A with 290-dimensional state, 52-dimensional action, 30 FPS, and different image sizes.
-
-No joint is fabricated or discarded to force a 14-dimensional representation. No frame resampling, image resizing, or action reinterpretation is performed merely to imitate RoboDojo metadata. Models must use the declared G2A schema or an independently documented embodiment adapter.
-
-## HDF5 archive policy
-
-The preferred raw archive is **one HDF5 file per episode**, organized by task and robot, with `instruction`, `state`, `action`, `vision`, and `additional_info` groups and seven-digit episode filenames, consistent with RoboDojo's raw-data organization.
-
-The currently verified input is already a LeRobot export. An HDF5 reconstructed from it will be explicitly labeled **a derived HDF5 archival export**, not an original raw sensor recording. Original unrecorded precision, depth, calibration, or sensors cannot be recovered from a compressed export and will not be fabricated.
-
-All available numeric values and source field/index mappings must be retained. Source-provided camera calibration may be copied only with its documented coordinate-frame meaning. Image encoding and any transformation must be declared; decoded RGB must be archived losslessly unless the original encoded image bytes are available. This may differ from the JPEG payload of RoboDojo, and must not be described as byte-identical raw data.
-
-## Transfer and verification
-
-1. Keep the verified source read-only; write each export to a new staging directory.
-2. Export and validate pilot task `16249`, then freeze the export with SHA-256 checksums.
-3. Validate with an official LeRobot v3.0-compatible loader, compare every numeric frame against the source, check all copied-video hashes, and decode representative first/last frames from every episode.
-4. Upload only the frozen pilot export to `data/ValueModelBenchmark_lerobot_v30_video/`. Do not upload private metadata attachments, credentials, internal paths, or operational logs.
-5. Keep the Hugging Face upload cache for resumability. Verify the remote inventory and hashes, download the complete pilot into a separate verification directory, and run the official loader against that downloaded copy before declaring the release ready.
-6. Expand to the remaining 19 tasks only after pilot remote verification passes. The full training split and counts must be updated to describe the actual uploaded release, never the planned one.
-
-The initial source-to-Mac copy and pilot remote verification have passed. The full training release has been uploaded and made public; its one-time final remote verification has passed, including all numeric rows, episode boundaries, 6,000 byte-identical videos and first/middle/last frames of every episode.
-
-## Download and loading
-
-公开仓库可匿名下载。这是多子集容器仓库，LeRobot 的 `root` 必须指向实际数据子目录，而不是仓库根目录。示例固定已上传的数据 commit 以便复现，**该数据版本已通过全量远端验收**。
-
-Anonymous download is supported. Pass the actual training subdirectory, not the repository root, to the LeRobot loader. The example pins the uploaded data commit for reproducibility; **this data revision passed full-release remote acceptance**.
+LeRobot 的 `root` 指向所选子集目录。The LeRobot `root` must point to the selected subset directory.
 
 ```python
 from pathlib import Path
@@ -230,14 +210,10 @@ from huggingface_hub import snapshot_download
 from lerobot.datasets.lerobot_dataset import LeRobotDataset
 
 repo_id = "wzlllong/value-model-benchmark-training-data"
-subset_id = "real-world-full-shot-data"
-prefix = "data/ValueModelBenchmark_lerobot_v30_video"
-revision = "9112657a3231cc1a969f572fe244b245d99a0a3b"
-
+prefix = "data/real-world/full-shot"
 snapshot = snapshot_download(
     repo_id=repo_id,
     repo_type="dataset",
-    revision=revision,
     allow_patterns=[f"{prefix}/**"],
     token=False,
 )
@@ -246,16 +222,11 @@ dataset = LeRobotDataset(
     root=Path(snapshot) / prefix,
     video_backend="pyav",
 )
-print(subset_id, dataset.num_episodes, dataset.num_frames)
-print(dataset[0]["action"].shape, dataset[0]["observation.state"].shape)
+sample = dataset[0]
+arm_joint_targets = sample["action"][16:30]
+end_effector_positions = sample["action"][2:8].reshape(2, 3)
+end_effector_orientations = sample["action"][8:16].reshape(2, 4)
+print(dataset.num_episodes, dataset.num_frames)
 ```
 
-Pilot and full-release loading were verified with official LeRobot 0.4.2 and PyAV 15.1.0. Expected metadata: 2,000 episodes, 1,530,987 frames, action shape `(52,)`, state shape `(290,)`. Pilot 和全量数据使用官方 LeRobot 0.4.2 与 PyAV 15.1.0 验证读取；预期规模与维度如上，完整数据验收状态见发布说明。
-
-## Reference
-
-Reference repository: [RoboDojo](https://huggingface.co/datasets/RoboDojo-Benchmark/RoboDojo)
-
-Reference snapshot: `35efbc7dedfdbeeb6e95fb749bd885d73d483e41`, inspected on 2026-10-14.
-
-LeRobot dataset v3.0 documentation: [LeRobot v3.0](https://huggingface.co/docs/lerobot/lerobot-dataset-v3)
+[LeRobot dataset v3.0 documentation / 格式说明](https://huggingface.co/docs/lerobot/lerobot-dataset-v3)

data/ValueModelBenchmark_lerobot_v30_video/README.md

@@ -1,28 +0,0 @@
----
-tags:
-- lerobot
-- robotics
----
-# G2A real-robot training data
-
-LeRobot v3.0 export. Preserve 52-dimensional actions, 290-dimensional states,
-original numerical values (except global episode/frame/task indices), timestamps,
-30 fps, resolutions, and all three HEVC camera streams without re-encoding.
-Field descriptions in `meta/info.json` retain original units, coordinate frames,
-and XYZW quaternion ordering. Gripper values are not normalized.
-
-Camera aliases are recorded in `provenance/conversion.json`; image metadata is
-CHW, matching decoded tensors. This is format compatibility, not compatibility
-with RoboDojo's action space or its policies.
-
-Tabular statistics and global quantiles are exact. RGB image statistics use
-documented temporal/spatial sampling and omit image quantiles.
-
-Original dataset/task/episode indices appear in the provenance mapping.
-Unreviewed Agibot auxiliary sidecars, user identifiers, robot serials, internal
-storage URIs, endpoints, and logs are not included. The source remains unchanged.
-This does not assert coverage of any separately reported original 1.3 TB assets.
-
-Private pilot authorization is recorded in provenance. Public release and license
-remain pending. This document belongs inside the training-version subdirectory;
-the repository-root dataset card is maintained separately.

data/ValueModelBenchmark_lerobot_v30_video/provenance/build_complete.json

@@ -1,6 +0,0 @@
-{
-  "episodes": 2000,
-  "frames": 1530987,
-  "tasks": 20,
-  "publication_approved": false
-}

data/real-world/full-shot/README.md

@@ -0,0 +1,36 @@
+---
+language:
+- en
+- zh
+tags:
+- robotics
+- lerobot
+- real-world
+- full-shot
+---
+
+# Real-World Full-Shot Training Data / 真机 Full-Shot 训练集
+
+**Subset ID / 子集标识：`real-world-full-shot-data`**
+
+## 中文
+
+本子集包含 G2A 机器人的 20 个真实世界任务、2,000 条轨迹，每个任务 100 条；共 1,530,987 个时间步，采样率 30 FPS，总轨迹时长约 14.18 小时。
+
+使用 LeRobot v3.0 格式，包含 290 维状态、52 维动作，以及头部、左腕、右腕三路 RGB HEVC 视频。头部图像为 640 × 400，左右腕各为 640 × 528。
+
+**动作同时包含关节目标与末端目标位姿。**双臂关节目标为 `action[16:30]`；双臂末端目标位置为 `action[2:8]`，目标朝向为 `action[8:16]`。末端目标由关节 action 通过正运动学计算；位置采用 `base_link` 坐标系，单位 m，朝向为 XYZW 四元数。完整索引和字段解释见 `meta/info.json`。
+
+原始数值与视频编码保持不变，夹爪值未归一化。`meta/episodes/` 定义轨迹边界，`provenance/episode_mapping.json` 记录源任务与轨迹映射，`checksums.sha256` 用于文件完整性检查。Full-shot 表示每个任务的完整样本集合，不代表成功标签或额外 reward / value target。数据许可尚未指定。
+
+## English
+
+This subset contains 2,000 G2A real-world episodes across 20 tasks, with 100 episodes per task: 1,530,987 trajectory time steps at 30 FPS, totaling approximately 14.18 hours.
+
+The LeRobot v3.0 dataset includes 290-dimensional states, 52-dimensional actions, and head/left-wrist/right-wrist RGB HEVC videos. Head resolution is 640 × 400; each wrist resolution is 640 × 528.
+
+**Actions contain both joint targets and end-effector target poses.**Arm joint targets are `action[16:30]`; arm end-effector target positions are `action[2:8]`, and orientations are `action[8:16]`. End-effector targets are computed from joint actions by forward kinematics. Positions use `base_link` coordinates in m; orientations use XYZW quaternions. Full field/index definitions are in `meta/info.json`.
+
+Original numeric values and video encoding are retained; gripper values are not normalized. `meta/episodes/` defines episode boundaries, `provenance/episode_mapping.json` records source mappings, and `checksums.sha256` supports integrity checks. Full-shot denotes the complete episode collection per task, not success labels or additional reward/value targets. No data license has been specified.
+
+[Dataset overview and task inventory / 数据集总览与任务清单](../../../README.md)

data/{ValueModelBenchmark_lerobot_v30_video → real-world/full-shot}/checksums.sha256

@@ -1,4 +1,4 @@
-87e25808c8418c9caec5daa81f1d6a4a8e66470ee864a038930144a37f5777e0  README.md
+910903dd1c480bb40be137071a0d64bd2a9ff1ee22659ddb8c3055a01b8325a1  README.md
 80e785f6a7f703e48e25f00ee08ccb7a9d1698960fecbfc0f383d031df184959  data/chunk-000/file-000.parquet
 4a1370a74adc71fa8725f7b796b09c488d1a9894377f4e2b0c59d84460c9b147  data/chunk-000/file-001.parquet
 233868f5d7867dc47ce76f5788c67d8bee0bafe9f7f5a6c941dc210c23509a85  data/chunk-000/file-002.parquet
@@ -2003,8 +2003,7 @@ a4a93e2e867766d5198c272f8498c64c7f016fb6e1b5251d78630ec9f909c023  data/chunk-001
 513d8ce394df7704c6a0b1ce9164c6afc3d76cc6493d55f513420ad06b528c46  meta/info.json
 8a32ce3c5053c6063ecdeb2f8933c1a398ac9e6d243723fa1d3347b15bb199bb  meta/stats.json
 fe80e4c17ad7b530ffc48b089fe461d1dbb11d77b5507d0c921832f3d337d855  meta/tasks.parquet
-a0ee851751b0fac728f5688040296cacf27c32ae679e6e13ffa64009e414e8ed  provenance/build_complete.json
-e3e83f1c2700fee38274893a3c71853c305c6d90436d897dd95a78dce21c39f9  provenance/conversion.json
+e2e0bf4eb6db0cecce71cb95b92560b2d0f56e55af3d263a71859e3aa06cd0e7  provenance/conversion.json
 fd06b7d6987fcc62e607c5ca9b61ee73dc506dd1317be453bb96d444d55d65ee  provenance/episode_mapping.json
 bc5ffa087bf5be7e648e16d3f2135bec43ba6fa263ab07e62e89a20cbfa3c0bf  provenance/episodes.parquet
 faa60b6cb0e015a9de63bcdfb17580126193730a1adc26276d6b35ee70f799ac  videos/observation.images.cam_high/chunk-000/file-000.mp4