Refine RoboValue dataset presentation and locally verified quick start

Document ARX and AgiBot Genie02, reviewed and recollected simulation sources, common trajectory contents, and unified displayed simulation task names. Align bilingual card layout with the RoboValue project README and provide a locally executed LeRobot 0.4.2 example. Training data and provenance remain unchanged.

README.md

@@ -15,46 +15,71 @@ tags:
 - one-shot
 ---
 
-# RoboValue Training Data
+<div align="center">
+  <img src="https://huggingface.co/datasets/wzlllong/value-model-benchmark-training-data/resolve/main/assets/robovalue-logo.png" alt="RoboValue" width="420" />
+  <h1>Training Data for Robotic Value Models</h1>
+  <p><strong>Simulation &amp; Real-World · Full-Shot &amp; One-Shot</strong></p>
+  <p>
+    <a href="https://github.com/RoboValue-Benchmark/RoboValue"><img src="https://img.shields.io/badge/GitHub-RoboValue-24292F?logo=github" alt="RoboValue GitHub" /></a>
+    <a href="https://robovalue-benchmark.github.io/doc/"><img src="https://img.shields.io/badge/Documentation-8A2BE2?logo=readthedocs&amp;logoColor=white" alt="Documentation" /></a>
+    <a href="https://robovalue-benchmark.github.io/"><img src="https://img.shields.io/badge/Website-527BC3?logo=googlechrome&amp;logoColor=white" alt="Website" /></a>
+  </p>
+  <p><strong>English</strong> · <a href="https://huggingface.co/datasets/wzlllong/value-model-benchmark-training-data/blob/main/README_zh.md">简体中文</a></p>
+</div>
 
-**English** · [简体中文](https://huggingface.co/datasets/wzlllong/value-model-benchmark-training-data/blob/main/README_zh.md)
+RoboValue provides expert demonstrations for **robotic value models**, covering **15 simulation tasks** and **20 real-world tasks**. Both domains offer Full-Shot and One-Shot subsets in the shared **LeRobot Dataset v3.0** format.
 
-Expert demonstrations for **robotic value models**, covering simulated and real-world manipulation tasks. Each domain has two demonstration-access settings: **Full-Shot** for training or fine-tuning, and **One-Shot** for a single task-specific reference or adaptation example.
+<p align="center">
+  <a href="#overview">Overview</a> · <a href="#simulation">Simulation</a> · <a href="#real-world">Real-World</a> · <a href="#repository-layout">Repository</a> · <a href="#download-and-reading">Download &amp; Read</a>
+</p>
 
-All subsets use **LeRobot Dataset v3.0**, with per-step Parquet data, three RGB camera streams, and task, episode, and statistics metadata.
+## 1. Dataset Overview <a name="overview"></a>
 
-**Dataset size:** Simulation: **1,500 Full-Shot episodes** and **15 One-Shot episodes** across 15 tasks. Real-World: **2,000 Full-Shot episodes** and **20 One-Shot episodes** across 20 tasks. The two One-Shot subsets contain **35 episodes, 23,747 synchronized rows, and 105 original camera videos** in total.
+| Domain | Platform | Tasks | Full-Shot | One-Shot |
+| --- | --- | ---: | ---: | ---: |
+| [**Simulation**](#simulation) | ARX | 15 | **1,500 episodes** | **15 episodes** |
+| [**Real-World**](#real-world) | AgiBot Genie02 | 20 | **2,000 episodes** | **20 episodes** |
 
-[Simulation](#2-simulation) · [Real-world data](#3-real-world) · [Quick start](#5-download-and-reading) · [Repository files](#4-repository-layout)
+| Setting | Demonstrations | Use |
+| --- | --- | --- |
+| **Full-Shot** | 100 per task | Training or fine-tuning on the complete expert collection |
+| **One-Shot** | 1 per task | Reference conditioning or task-specific adaptation |
 
-## 1. Dataset Overview
+The two One-Shot subsets contain **35 episodes, 23,747 time steps, and 105 camera videos** in total. These are standard in-domain expert demonstrations; value or progress supervision is constructed by the training method.
 
-| Domain | Tasks | Full-Shot | One-Shot |
-| --- | ---: | ---: | ---: |
-| [Simulation](#2-simulation) | 15 | **1,500** | **15** |
-| [Real-World](#3-real-world) | 20 | **2,000** | **20** |
+## 2. Simulation <a name="simulation"></a>
 
-The table counts **episodes**. Full-Shot contains **100 episodes per task**, and One-Shot contains **one episode per task**. One-Shot task–episode mappings are recorded in each subset's `provenance/episode_mapping.json`.
+### 2.1 Overview
+
+Simulation demonstrations use the **ARX** robot platform. They are based on the open-source [**RoboDojo**](https://huggingface.co/datasets/RoboDojo-Benchmark/RoboDojo) dataset, reviewed by our team with selected demonstrations recollected. They cover **15 manipulation tasks** at **25 FPS**.
 
-**Which setting should I use?**
+**What does one trajectory contain?**
 
-- **Full-Shot:** learn or fine-tune task-conditioned value/progress representations using the full demonstration collection.
-- **One-Shot:** condition a model on one reference demonstration, or perform task-specific adaptation under a one-example budget.
+In both Full-Shot and One-Shot, an **episode** is a complete task demonstration containing the task instruction, per-step states and actions, identifiers and timestamps, and three RGB camera streams. Episode metadata records its length, data-row range, and video segments for each camera.
 
-The dataset provides **standard in-domain expert demonstrations** for robotic value models. Fields include robot states, actions, task and episode indices, timestamps, and three RGB views. Value or progress supervision is constructed by the training method.
+| Content | Representation |
+| --- | --- |
+| Robot state | `observation.state`: `float32[14]` |
+| Robot action | `action`: `float32[14]` |
+| RGB observations | High, left-wrist, and right-wrist views, each **640 × 480** |
+| Instruction and identity | Task instruction, `task_index`, `episode_index`, `frame_index` |
+| Time | `timestamp` in seconds; **25 FPS** |
 
-## 2. Simulation
+<details>
+<summary><strong>Simulation feature reference: states, actions, and cameras</strong></summary>
 
-### 2.1 Overview
+State and action share the same 14-dimensional field order. Python slices are stop-exclusive.
 
-The simulation domain contains **15 manipulation tasks**, organized into Full-Shot and One-Shot subsets. Robot type is `unified_robot`, sampling rate is **25 FPS**, and all three RGB cameras have **640 × 480** resolution. State and action are both `float32[14]`.
+| Slice | Source field names |
+| --- | --- |
+| `[0:7]` | `left_joint_0` … `left_joint_6` |
+| `[7:14]` | `right_joint_0` … `right_joint_6` |
+
+Camera keys are `observation.images.cam_high`, `observation.images.cam_left_wrist`, and `observation.images.cam_right_wrist`. Loaded images have shape **[3, 480, 640]** (CHW).
 
-| Setting | Per task | Episodes | Data |
-| --- | ---: | ---: | --- |
-| [**Full-Shot**](#22-full-shot) | 100 | **1,500** | [Files](https://huggingface.co/datasets/wzlllong/value-model-benchmark-training-data/tree/main/data/simulation/full-shot) |
-| [**One-Shot**](#23-one-shot) | 1 | **15** | [Files](https://huggingface.co/datasets/wzlllong/value-model-benchmark-training-data/tree/main/data/simulation/one-shot) |
+[Full-Shot schema](https://huggingface.co/datasets/wzlllong/value-model-benchmark-training-data/blob/main/data/simulation/full-shot/meta/info.json) · [One-Shot schema](https://huggingface.co/datasets/wzlllong/value-model-benchmark-training-data/blob/main/data/simulation/one-shot/meta/info.json)
 
-The catalog below uses Full-Shot source names. One-Shot retains `sweep_objects` for the corresponding Full-Shot `sweep_blocks` task; both instructions describe sweeping blocks with a broom and dustpan.
+</details>
 
 <details>
 <summary><strong>Simulation task catalog</strong></summary>
@@ -81,104 +106,49 @@ The catalog below uses Full-Shot source names. One-Shot retains `sweep_objects`
 
 ### 2.2 Full-Shot
 
-**15 tasks × 100 demonstrations = 1,500 episodes and 789,612 time steps.**
+**15 tasks × 100 demonstrations = 1,500 episodes.**
 
-| Item | Value |
-| --- | --- |
-| Synchronized data rows | **789,612** at **25 FPS** |
-| State / action | `float32[14]` / `float32[14]` |
-| Robot type | `unified_robot` |
-| Cameras | Three RGB views, **640×480**, **AV1 / H.264**; **4,500 videos** |
-| Storage / duration | **≈45.47 GB**; **≈8.77 hours**, ≈21.06 seconds per episode |
-| Split | `train: 0:1500` |
+| Time steps | Videos | Size | Total duration | Training split |
+| ---: | --- | ---: | ---: | --- |
+| **789,612** | 4,500 · AV1 / H.264 | ≈45.47 GB | ≈8.77 hours | `train: 0:1500` |
 
-Sources are **1,393 episodes** from [RoboDojo](https://huggingface.co/datasets/RoboDojo-Benchmark/RoboDojo) and **107 supplemental episodes**. Source collections and episode indices are recorded in `provenance/episodes.parquet`.
+[**Browse files →**](https://huggingface.co/datasets/wzlllong/value-model-benchmark-training-data/tree/main/data/simulation/full-shot) · [Download and read](#download-and-reading)
 
 ### 2.3 One-Shot
 
-**15 tasks × 1 demonstration = 15 episodes and 8,080 time steps.** Sampling rate is **25 FPS**, and state and action are both `float32[14]`. Three **640 × 480** RGB cameras provide **45 AV1 videos**. Split: `train: 0:15`; size: **≈0.50 GB**.
+**15 tasks × 1 demonstration = 15 episodes.** Uses the same task definitions as Full-Shot.
 
-Source task name `sweep_objects` corresponds to Full-Shot `sweep_blocks`; the instruction uses a broom and dustpan to sweep blocks. The [task–episode–row mapping](https://huggingface.co/datasets/wzlllong/value-model-benchmark-training-data/blob/main/data/simulation/one-shot/provenance/episode_mapping.json) contains source task names, episode indices, and data-row boundaries.
+| Time steps | Videos | Size | Total duration | Training split |
+| ---: | --- | ---: | ---: | --- |
+| **8,080** | 45 · AV1 | ≈0.50 GB | ≈5.39 minutes | `train: 0:15` |
 
-Read prefix: `data/simulation/one-shot`; scale: **15 episodes and 8,080 time steps**.
+[**Browse files →**](https://huggingface.co/datasets/wzlllong/value-model-benchmark-training-data/tree/main/data/simulation/one-shot) · [Task–episode–row mapping](https://huggingface.co/datasets/wzlllong/value-model-benchmark-training-data/blob/main/data/simulation/one-shot/provenance/episode_mapping.json) · [Download and read](#download-and-reading)
 
-## 3. Real-World
+## 3. Real-World <a name="real-world"></a>
 
 ### 3.1 Overview
 
-Real robot demonstrations are collected with **G2A** and cover **20 manipulation tasks**, including object placement and sorting, household organization, drawers, repeated presses, kitchen activities, and packaging.
-
-| Setting | Per task | Episodes | Data |
-| --- | ---: | ---: | --- |
-| [**Full-Shot**](#32-full-shot) | 100 | **2,000** | [Files](https://huggingface.co/datasets/wzlllong/value-model-benchmark-training-data/tree/main/data/real-world/full-shot) |
-| [**One-Shot**](#33-one-shot) | 1 | **20** | [Files](https://huggingface.co/datasets/wzlllong/value-model-benchmark-training-data/tree/main/data/real-world/one-shot) |
-
-<details>
-<summary><strong>Task catalog: all 20 tasks and short descriptions</strong></summary>
-
-Indices match the training data's `task_index`. Display names and descriptions follow Appendix D of the RoboValue paper; the original task strings in `meta/tasks.parquet` are retained.
-
-| Index | Task and description |
-| --- | --- |
-| 0 | **Transfer Blocks Using Plate**<br>Load red, green, then blue blocks onto a plate and pour them into the box. |
-| 1 | **Enter Keypad Number Sequence**<br>Enter 2-8-2-6-8, releasing and returning between presses. |
-| 2 | **Set the Table**<br>Take the plate, bowl, fork and knife from the crate and set the table. |
-| 3 | **Arrange Seasoning Jars**<br>Place three seasoning jars on the two-tier rack. |
-| 4 | **Load Plastic Plates Into Rack**<br>Insert three plastic plates into the rack. |
-| 5 | **Press Button Five Times**<br>Press the button five times, releasing and returning after each press. |
-| 6 | **Place Pot On Induction Cooker**<br>Place the pot on the cooker, store the spatula and cover the pot. |
-| 7 | **Organize Desktop Workstation**<br>Store the pen and eraser, close the drawer, switch off and unplug the lamp. |
-| 8 | **Organize Books And Pen**<br>Align one book; place the pen in the other, close it and stack it. |
-| 9 | **Arrange Family Tea Set**<br>Arrange the cup, teapot/lid, canister and tongs in their target areas. |
-| 10 | **Operate Drawers In Sequence**<br>Open the drawers bottom-to-top, then close them top-to-bottom. |
-| 11 | **Sort Blocks in Order**<br>Place red, yellow and blue blocks on their designated drawer units, in order. |
-| 12 | **Fill And Place Pen Holder**<br>Insert four pens into the holder and place it upright. |
-| 13 | **Organize Cosmetics Storage Box**<br>Store the lipstick and eyebrow pencil, close the drawer, then place the serum and cleansing oil. |
-| 14 | **Strike Drum Five Times**<br>Pick up the drumstick, strike five times and return it. |
-| 15 | **Apply Dish Soap Repeatedly**<br>Hold the cloth beneath the outlet and pump dish soap five times. |
-| 16 | **Season Meat Skewers Repeatedly**<br>Season the meat skewers with the shaker five times. |
-| 17 | **Pack Items For Shipping**<br>Pack the mouse, then the wired earphones, and cover both with bubble wrap. |
-| 18 | **Arrange Flowers and Reed Diffuser**<br>Arrange three flowers, place the diffuser bottle and insert the reed. |
-| 19 | **Prepare Breakfast**<br>Place the buns in their designated positions, add a steamer tier and cover it. |
-
-</details>
-
-### 3.2 Full-Shot
-
-**20 tasks × 100 demonstrations = 2,000 episodes and 1,530,987 time steps.**
-
-[Download and read](#5-download-and-reading) · [Browse files](https://huggingface.co/datasets/wzlllong/value-model-benchmark-training-data/tree/main/data/real-world/full-shot) · [Feature definitions](https://huggingface.co/datasets/wzlllong/value-model-benchmark-training-data/blob/main/data/real-world/full-shot/meta/info.json)
-
-**Key statistics**
-
-| Item | Value |
-| --- | --- |
-| Coverage | **20 tasks × 100 episodes = 2,000 episodes** |
-| Time steps / sampling | **1,530,987** at **30 FPS** |
-| Trajectory duration | **≈14.18 hours** total; **≈25.52 seconds** per episode on average |
-| Cameras | **3 RGB views:** head, left wrist, right wrist |
-| Resolution | Head **640 × 400**; each wrist **640 × 528** |
-| Storage | **≈23.23 GB**; Parquet + HEVC video |
-| Split | `train: 0:2000` |
-
-Duration and time steps are counted once per episode.
+Real-world demonstrations are collected with the **AgiBot Genie02** robot. They cover **20 manipulation tasks**, including object placement and sorting, household organization, drawers, repeated presses, kitchen activities, and packaging, at **30 FPS**.
 
 **What does one trajectory contain?**
 
-An **episode** is a complete task demonstration containing task/episode identifiers, per-step robot observations and actions, and three camera streams. Episode metadata records length, data-file references, data-row boundaries, and camera-video segments.
+In both Full-Shot and One-Shot, an **episode** is a complete task demonstration containing the task instruction, per-step robot observations and actions, identifiers and timestamps, and RGB videos from the head, left wrist, and right wrist. Episode metadata records its length, data-row range, and video segments for each camera.
 
-| Signal | Representation |
+| Content | Representation |
 | --- | --- |
-| `action` | **52 float32 values:** joint targets **and** end-effector target poses, plus gripper, head, waist, and base-control fields |
-| `observation.state` | **290 float32 values:** joint and end-effector state, including position, orientation, velocity, torque/force, and auxiliary fields |
-| RGB observations | Head, left-wrist, and right-wrist images; video features use CHW metadata |
-| Identity and time | `episode_index`, `task_index`, `frame_index`, and per-step `timestamp` in seconds |
+| Robot state | `observation.state`: `float32[290]`, including joint/end-effector state and auxiliary fields |
+| Robot action | `action`: `float32[52]`, including joint targets, end-effector target poses, and other control fields |
+| RGB observations | Head **640 × 400**; each wrist **640 × 528** |
+| Instruction and identity | Task instruction, `task_index`, `episode_index`, `frame_index` |
+| Time | `timestamp` in seconds; **30 FPS** |
 
-**End-effector action targets are computed by forward kinematics from joint actions.** Positions use meters in `base_link`, and orientations use XYZW quaternions. Grippers use their original opening or stroke values. The 52-dimensional action vector includes end-effector poses, joint targets, and other control fields; slices are listed below.
+Camera keys are `observation.images.cam_high`, `observation.images.cam_left_wrist`, and `observation.images.cam_right_wrist`. Loaded images use **CHW** order.
 
 <details>
 <summary><strong>Action reference: all 52 dimensions</strong></summary>
 
+**End-effector action targets are computed by forward kinematics from joint actions.** Positions use meters in `base_link`, and orientations use XYZW quaternions. Grippers use their original opening or stroke values. The 52-dimensional action vector includes end-effector poses, joint targets, and other control fields; slices are listed below.
+
 Python slices are stop-exclusive.
 
 | Slice | Meaning |
@@ -218,88 +188,136 @@ The table summarizes the main fields of the 290-dimensional state vector. Comple
 
 </details>
 
+[Full-Shot schema](https://huggingface.co/datasets/wzlllong/value-model-benchmark-training-data/blob/main/data/real-world/full-shot/meta/info.json) · [One-Shot schema](https://huggingface.co/datasets/wzlllong/value-model-benchmark-training-data/blob/main/data/real-world/one-shot/meta/info.json)
+
+<details>
+<summary><strong>Task catalog: all 20 tasks and short descriptions</strong></summary>
+
+Indices match the training data's `task_index`. Display names and descriptions follow Appendix D of the RoboValue paper; the original task strings in `meta/tasks.parquet` are retained.
+
+| Index | Task and description |
+| --- | --- |
+| 0 | **Transfer Blocks Using Plate**<br>Load red, green, then blue blocks onto a plate and pour them into the box. |
+| 1 | **Enter Keypad Number Sequence**<br>Enter 2-8-2-6-8, releasing and returning between presses. |
+| 2 | **Set the Table**<br>Take the plate, bowl, fork and knife from the crate and set the table. |
+| 3 | **Arrange Seasoning Jars**<br>Place three seasoning jars on the two-tier rack. |
+| 4 | **Load Plastic Plates Into Rack**<br>Insert three plastic plates into the rack. |
+| 5 | **Press Button Five Times**<br>Press the button five times, releasing and returning after each press. |
+| 6 | **Place Pot On Induction Cooker**<br>Place the pot on the cooker, store the spatula and cover the pot. |
+| 7 | **Organize Desktop Workstation**<br>Store the pen and eraser, close the drawer, switch off and unplug the lamp. |
+| 8 | **Organize Books And Pen**<br>Align one book; place the pen in the other, close it and stack it. |
+| 9 | **Arrange Family Tea Set**<br>Arrange the cup, teapot/lid, canister and tongs in their target areas. |
+| 10 | **Operate Drawers In Sequence**<br>Open the drawers bottom-to-top, then close them top-to-bottom. |
+| 11 | **Sort Blocks in Order**<br>Place red, yellow and blue blocks on their designated drawer units, in order. |
+| 12 | **Fill And Place Pen Holder**<br>Insert four pens into the holder and place it upright. |
+| 13 | **Organize Cosmetics Storage Box**<br>Store the lipstick and eyebrow pencil, close the drawer, then place the serum and cleansing oil. |
+| 14 | **Strike Drum Five Times**<br>Pick up the drumstick, strike five times and return it. |
+| 15 | **Apply Dish Soap Repeatedly**<br>Hold the cloth beneath the outlet and pump dish soap five times. |
+| 16 | **Season Meat Skewers Repeatedly**<br>Season the meat skewers with the shaker five times. |
+| 17 | **Pack Items For Shipping**<br>Pack the mouse, then the wired earphones, and cover both with bubble wrap. |
+| 18 | **Arrange Flowers and Reed Diffuser**<br>Arrange three flowers, place the diffuser bottle and insert the reed. |
+| 19 | **Prepare Breakfast**<br>Place the buns in their designated positions, add a steamer tier and cover it. |
+
+</details>
+
+### 3.2 Full-Shot
+
+**20 tasks × 100 demonstrations = 2,000 episodes.**
+
+| Time steps | Videos | Size | Total duration | Training split |
+| ---: | --- | ---: | ---: | --- |
+| **1,530,987** | 6,000 · HEVC | ≈23.23 GB | ≈14.18 hours | `train: 0:2000` |
+
+[**Browse files →**](https://huggingface.co/datasets/wzlllong/value-model-benchmark-training-data/tree/main/data/real-world/full-shot) · [Download and read](#download-and-reading)
+
 ### 3.3 One-Shot
 
-**20 tasks × 1 demonstration = 20 episodes and 15,667 time steps.** Sampling rate is **30 FPS**, state is `float32[290]`, and action is `float32[52]`. Three RGB cameras provide **60 HEVC (H.265) videos**. Split: `train: 0:20`; size: **≈0.23 GB**.
+**20 tasks × 1 demonstration = 20 episodes.** Uses the same task definitions as Full-Shot.
 
-The [task–episode–row mapping](https://huggingface.co/datasets/wzlllong/value-model-benchmark-training-data/blob/main/data/real-world/one-shot/provenance/episode_mapping.json) contains source task identifiers, episode indices, and data-row boundaries. Camera fields are `cam_left_wrist`, `cam_right_wrist`, and `cam_high`, corresponding to left wrist, right wrist, and head views.
+| Time steps | Videos | Size | Total duration | Training split |
+| ---: | --- | ---: | ---: | --- |
+| **15,667** | 60 · HEVC | ≈0.23 GB | ≈8.70 minutes | `train: 0:20` |
 
-Read prefix: `data/real-world/one-shot`; scale: **20 episodes and 15,667 time steps**.
+[**Browse files →**](https://huggingface.co/datasets/wzlllong/value-model-benchmark-training-data/tree/main/data/real-world/one-shot) · [Task–episode–row mapping](https://huggingface.co/datasets/wzlllong/value-model-benchmark-training-data/blob/main/data/real-world/one-shot/provenance/episode_mapping.json) · [Download and read](#download-and-reading)
 
-## 4. Repository Layout
+## 4. Repository Layout <a name="repository-layout"></a>
 
 ```text
 README.md                         English dataset card
 README_zh.md                      Chinese dataset card
-subset_catalog.json               Machine-readable subset index
+subset_catalog.json               Subset index
 data/
 ├── simulation/
-│   ├── full-shot/                LeRobot dataset root
-│   └── one-shot/                 LeRobot dataset root
+│   ├── full-shot/                1,500 episodes · 15 tasks
+│   └── one-shot/                 15 episodes · 1 per task
 └── real-world/
-    ├── full-shot/                LeRobot dataset root
-    └── one-shot/                 LeRobot dataset root
+    ├── full-shot/                2,000 episodes · 20 tasks
+    └── one-shot/                 20 episodes · 1 per task
 ```
 
-Each subset directory serves as its own LeRobot `root`; all four subsets use the layout above.
+Each subset directory is an independent **LeRobot `root`**, containing per-step Parquet data, three original camera streams, and metadata.
 
 <details>
-<summary><strong>Format reference: files inside any released subset</strong></summary>
+<summary><strong>Subset files and metadata</strong></summary>
 
 | File or directory | Purpose |
 | --- | --- |
-| `data/chunk-*/file-*.parquet` | Per-step state, action, identifiers, and timestamps |
+| `data/chunk-*/file-*.parquet` | Per-step states, actions, identifiers, and timestamps |
 | `videos/<camera>/chunk-*/file-*.mp4` | RGB camera video streams |
-| `meta/info.json` | v3.0 version, features, shapes, sampling rate, and path templates |
-| `meta/tasks.parquet` | Original task strings and their task indices |
-| `meta/episodes/chunk-*/file-*.parquet` | Episode length, task, data boundaries, video offsets, and per-episode statistics |
+| `meta/info.json` | Format version, features, dimensions, FPS, and path templates |
+| `meta/tasks.parquet` | Task instructions and indices |
+| `meta/episodes/chunk-*/file-*.parquet` | Episode length, task, data-row boundaries, video offsets, and episode statistics |
 | `meta/stats.json` | Dataset-level feature statistics |
 | `provenance/` | Episode sources and task correspondence |
 | `checksums.sha256` | File integrity manifest |
 
-Camera feature keys are `observation.images.cam_high`, `observation.images.cam_left_wrist`, and `observation.images.cam_right_wrist`. Episode data boundaries are `dataset_from_index` and `dataset_to_index` (stop-exclusive); video timestamps are stored separately for each camera.
+Data rows span `[dataset_from_index, dataset_to_index)`; video segments are recorded separately for each camera. Time steps and duration are counted once per episode.
 
-[Official LeRobot Dataset v3.0 format guide](https://huggingface.co/docs/lerobot/lerobot-dataset-v3) · [Subset catalog](https://huggingface.co/datasets/wzlllong/value-model-benchmark-training-data/blob/main/subset_catalog.json)
+[Official LeRobot v3.0 format guide](https://huggingface.co/docs/lerobot/lerobot-dataset-v3) · [Machine-readable subset catalog](https://huggingface.co/datasets/wzlllong/value-model-benchmark-training-data/blob/main/subset_catalog.json)
 
 </details>
 
-## 5. Download and Reading
+## 5. Download and Reading <a name="download-and-reading"></a>
 
 ### 5.1 Choose a subset
 
-All four prefixes below are independent LeRobot roots. Storage uses decimal GB and includes the original videos.
+Select one `prefix` to download only that subset. Sizes use decimal GB.
 
-| Subset | `prefix` | Episodes | Rows | Videos | FPS | State / action dimensions | Size |
-| --- | --- | ---: | ---: | ---: | ---: | --- | ---: |
-| Simulation Full-Shot | `data/simulation/full-shot` | 1,500 | 789,612 | 4,500 | 25 | 14 / 14 | ≈45.47 GB |
-| Simulation One-Shot | `data/simulation/one-shot` | 15 | 8,080 | 45 | 25 | 14 / 14 | ≈0.50 GB |
-| Real-World Full-Shot | `data/real-world/full-shot` | 2,000 | 1,530,987 | 6,000 | 30 | 290 / 52 | ≈23.23 GB |
-| Real-World One-Shot | `data/real-world/one-shot` | 20 | 15,667 | 60 | 30 | 290 / 52 | ≈0.23 GB |
+| Subset | `prefix` | Episodes / time steps | Size |
+| --- | --- | ---: | ---: |
+| Simulation Full-Shot | `data/simulation/full-shot` | 1,500 / 789,612 | ≈45.47 GB |
+| Simulation One-Shot | `data/simulation/one-shot` | 15 / 8,080 | ≈0.50 GB |
+| Real-World Full-Shot | `data/real-world/full-shot` | 2,000 / 1,530,987 | ≈23.23 GB |
+| Real-World One-Shot | `data/real-world/one-shot` | 20 / 15,667 | ≈0.23 GB |
 
 ### 5.2 Download and read
 
-**The example loads Simulation Full-Shot.** Set `prefix` to either One-Shot row for a smaller first download. Public downloads require no token.
+This example loads the smaller **Simulation One-Shot** subset and returns states, actions, and all three camera frames. To switch subsets, change `prefix` to the corresponding path in the table. Public downloads require no token.
 
-The reader environment uses **Python 3.10+**, **LeRobot 0.4.2**, and **PyAV 15.1.0**. The data format is **LeRobot Dataset v3.0**.
+Use **Python 3.10+**, **LeRobot 0.4.2**, and **PyAV 15.1.0**:
 
 ```bash
-pip install "lerobot==0.4.2" "av==15.1.0"
+python -m pip install "lerobot==0.4.2" "av==15.1.0"
 ```
 
 ```python
+import os
 from pathlib import Path
 
 from huggingface_hub import snapshot_download
 from lerobot.datasets.lerobot_dataset import LeRobotDataset
 
 repo_id = "wzlllong/value-model-benchmark-training-data"
-prefix = "data/simulation/full-shot"
+prefix = "data/simulation/one-shot"
 revision = "3a81d8f75444d57ecae1d604de0d604130ec60cd"
+local_dir = Path(os.environ.get("ROBOVALUE_LOCAL_DIR", "./robovalue-data"))
+
 snapshot = snapshot_download(
     repo_id=repo_id,
     repo_type="dataset",
     revision=revision,
     allow_patterns=[f"{prefix}/**"],
+    local_dir=local_dir,
     token=False,
 )
 dataset = LeRobotDataset(
@@ -309,16 +327,31 @@ dataset = LeRobotDataset(
     video_backend="pyav",
 )
 sample = dataset[0]
-print(dataset.num_episodes, dataset.num_frames)
-print(sample["action"].shape, sample["observation.state"].shape)
+print("episodes / frames:", dataset.num_episodes, dataset.num_frames)
+print("action / state:", tuple(sample["action"].shape),
+      tuple(sample["observation.state"].shape))
+for camera in dataset.meta.camera_keys:
+    print(camera, tuple(sample[camera].shape))
 ```
 
-The example returns **1,500 episodes and 789,612 time steps**, with `action` and `observation.state` shapes of **[14]**. `dataset[0]` returns the first time step's state, action, and corresponding camera frames.
+Output:
 
-Download and reading use the same fixed revision and only files under the selected prefix. Section 5.1 lists all four prefixes, episode counts, and row counts.
+```text
+episodes / frames: 15 8080
+action / state: (14,) (14,)
+observation.images.cam_high (3, 480, 640)
+observation.images.cam_left_wrist (3, 480, 640)
+observation.images.cam_right_wrist (3, 480, 640)
+```
+
+Files are saved under `./robovalue-data`. Set the `ROBOVALUE_LOCAL_DIR` environment variable to an existing download directory to reuse local files. Download and reading use the same fixed data revision.
 
 ### 5.3 Integrity checks
 
-Each root contains `checksums.sha256`. To verify a downloaded subset, run `shasum -a 256 -c checksums.sha256` from that subset directory. Source and task mappings are in `provenance/`; schema and statistics are in `meta/info.json` and `meta/stats.json`.
+Each subset contains `checksums.sha256`. Run this command from the subset directory:
+
+```bash
+shasum -a 256 -c checksums.sha256
+```
 
-**Publication authorization:** This dataset is publicly released with authorization. The RoboDojo-derived portion follows the Apache-2.0 license.
+**Publication authorization:** All data in this dataset were authorized before public release.

README_zh.md

@@ -1,43 +1,68 @@
-# RoboValue 训练数据集
+<div align="center">
+  <img src="https://huggingface.co/datasets/wzlllong/value-model-benchmark-training-data/resolve/main/assets/robovalue-logo.png" alt="RoboValue" width="420" />
+  <h1>机器人价值模型训练数据集</h1>
+  <p><strong>Simulation &amp; Real-World · Full-Shot &amp; One-Shot</strong></p>
+  <p>
+    <a href="https://github.com/RoboValue-Benchmark/RoboValue"><img src="https://img.shields.io/badge/GitHub-RoboValue-24292F?logo=github" alt="RoboValue GitHub" /></a>
+    <a href="https://robovalue-benchmark.github.io/doc/"><img src="https://img.shields.io/badge/Documentation-8A2BE2?logo=readthedocs&amp;logoColor=white" alt="Documentation" /></a>
+    <a href="https://robovalue-benchmark.github.io/"><img src="https://img.shields.io/badge/Website-527BC3?logo=googlechrome&amp;logoColor=white" alt="Website" /></a>
+  </p>
+  <p><a href="https://huggingface.co/datasets/wzlllong/value-model-benchmark-training-data">English</a> · <strong>简体中文</strong></p>
+</div>
 
-**简体中文** · [English](https://huggingface.co/datasets/wzlllong/value-model-benchmark-training-data)
+RoboValue 提供面向**机器人价值模型**的专家示例，覆盖 **15 个仿真任务**和 **20 个真实世界任务**。两个领域均提供 Full-Shot 与 One-Shot，采用统一的 **LeRobot Dataset v3.0** 格式。
 
-本仓库面向**机器人价值模型（value model）**，提供仿真与真实世界操作任务的专家示例。每个领域均分为两种示例使用设置：**Full-Shot** 用于完整示例训练或微调，**One-Shot** 用于单条任务参考或单示例适配。
+<p align="center">
+  <a href="#overview">数据总览</a> · <a href="#simulation">仿真世界</a> · <a href="#real-world">真实世界</a> · <a href="#repository-layout">仓库目录</a> · <a href="#download-and-reading">下载与读取</a>
+</p>
 
-各子集采用 **LeRobot Dataset v3.0** 格式，包含逐时间步 Parquet 数据、三路 RGB 相机视频，以及任务、轨迹和统计元数据。
+## 1. 数据总览 <a name="overview"></a>
 
-**数据规模：** 仿真领域：15 个任务，**1,500 条 Full-Shot** 与 **15 条 One-Shot**；真机领域：20 个任务，**2,000 条 Full-Shot** 与 **20 条 One-Shot**。两个 One-Shot 子集合计 **35 条轨迹、23,747 行同步数据、105 个原始相机视频**。
+| 领域 | 采集平台 | 任务数 | Full-Shot | One-Shot |
+| --- | --- | ---: | ---: | ---: |
+| [**仿真世界**](#simulation) | ARX | 15 | **1,500 条** | **15 条** |
+| [**真实世界**](#real-world) | AgiBot Genie02 | 20 | **2,000 条** | **20 条** |
 
-[仿真世界](#2-仿真世界) · [真实世界](#3-真实世界) · [快速开始](#5-下载与读取) · [仓库目录](#4-仓库目录)
+| 使用设置 | 示例数量 | 用途 |
+| --- | --- | --- |
+| **Full-Shot** | 每任务 100 条 | 使用完整专家示例进行训练或微调 |
+| **One-Shot** | 每任务 1 条 | 以单条任务示例进行参考条件输入或任务适配 |
 
-## 1. 数据总览
+两个 One-Shot 子集合计 **35 条轨迹、23,747 个时间步、105 个相机视频**。本数据集提供标准域内专家示例；价值或进度监督由训练方法构造。
 
-| 领域 | 任务数 | Full-Shot | One-Shot |
-| --- | ---: | ---: | ---: |
-| [仿真世界](#2-仿真世界) | 15 | **1,500** | **15** |
-| [真实世界](#3-真实世界) | 20 | **2,000** | **20** |
+## 2. 仿真世界 <a name="simulation"></a>
 
-上表按**轨迹条数**统计。Full-Shot **每任务 100 条**，One-Shot **每任务 1 条**。One-Shot 的任务—轨迹对应关系记录在 `provenance/episode_mapping.json` 中。
+### 2.1 总览
+
+仿真数据使用 **ARX** 机器人平台，基于 [**RoboDojo**](https://huggingface.co/datasets/RoboDojo-Benchmark/RoboDojo) 开源数据整理，经团队审核后，对部分示例重新采集。数据覆盖 **15 个操作任务**，采样率为 **25 FPS**。
 
-**什么时候使用哪一种设置？**
+**一条轨迹包含什么？**
 
-- **Full-Shot：**使用完整专家示例集合，训练或微调任务条件下的价值、进度等表征。
-- **One-Shot：**以一条任务示例作为模型参考条件，或在单示例预算下进行任务适配。
+Full-Shot 与 One-Shot 中，每条 **episode** 都是一条完整的任务示例，包含任务指令、逐时间步状态与动作、轨迹标识和时间戳，以及三路 RGB 相机视频。轨迹元数据记录长度、数据行范围及各相机的视频片段位置。
 
-本数据集提供机器人价值模型的**标准域内专家示例**。数据包含机器人状态、动作、任务与轨迹索引、时间戳及三路 RGB 图像；价值或进度监督由训练方法构造。
+| 内容 | 表示 |
+| --- | --- |
+| 机器人状态 | `observation.state`：`float32[14]` |
+| 机器人动作 | `action`：`float32[14]` |
+| RGB 观测 | 高位、左腕、右腕三路视角，均为 **640 × 480** |
+| 指令与标识 | 任务指令、`task_index`、`episode_index`、`frame_index` |
+| 时间 | `timestamp`，单位为秒；**25 FPS** |
 
-## 2. 仿真世界
+<details>
+<summary><strong>仿真字段参考：状态、动作与相机</strong></summary>
 
-### 2.1 总览
+状态与动作使用相同的 14 维字段排列，Python 切片右端不包含在内。
 
-仿真数据包含 **15 个操作任务**，分为 Full-Shot 与 One-Shot 两类。机器人类型为 `unified_robot`，采样率为 **25 FPS**，三路 RGB 图像分辨率为 **640 × 480**，状态与动作均为 `float32[14]`。
+| 切片 | 源字段名称 |
+| --- | --- |
+| `[0:7]` | `left_joint_0` … `left_joint_6` |
+| `[7:14]` | `right_joint_0` … `right_joint_6` |
+
+相机字段为 `observation.images.cam_high`、`observation.images.cam_left_wrist` 和 `observation.images.cam_right_wrist`。读取后的图像为 **[3, 480, 640]**（CHW）。
 
-| 设置 | 每任务 | 总轨迹数 | 数据 |
-| --- | ---: | ---: | --- |
-| [**Full-Shot**](#22-full-shot) | 100 | **1,500** | [文件](https://huggingface.co/datasets/wzlllong/value-model-benchmark-training-data/tree/main/data/simulation/full-shot) |
-| [**One-Shot**](#23-one-shot) | 1 | **15** | [文件](https://huggingface.co/datasets/wzlllong/value-model-benchmark-training-data/tree/main/data/simulation/one-shot) |
+[Full-Shot 字段定义](https://huggingface.co/datasets/wzlllong/value-model-benchmark-training-data/blob/main/data/simulation/full-shot/meta/info.json) · [One-Shot 字段定义](https://huggingface.co/datasets/wzlllong/value-model-benchmark-training-data/blob/main/data/simulation/one-shot/meta/info.json)
 
-下表使用 Full-Shot 的源任务名称。One-Shot 对应任务保留源名 `sweep_objects`，它对应 Full-Shot 的 `sweep_blocks`；两者指令均为用扫帚和簸箕扫积木。
+</details>
 
 <details>
 <summary><strong>仿真任务清单</strong></summary>
@@ -64,104 +89,49 @@
 
 ### 2.2 Full-Shot
 
-**15 个任务 × 100 条 = 1,500 条轨迹，共 789,612 个时间步。**
+**15 个任务 × 100 条 = 1,500 条轨迹。**
 
-| 项目 | 数值 |
-| --- | --- |
-| 同步数据行 / 采样率 | **789,612 / 25 FPS** |
-| 状态 / 动作 | `float32[14]` / `float32[14]` |
-| 机器人类型 | `unified_robot` |
-| 相机 | 三路 **640×480 AV1 / H.264** RGB，共 **4,500 个视频** |
-| 存储 / 时长 | **≈45.47 GB**；总时长 **≈8.77 小时**，每条平均 ≈21.06 秒 |
-| 划分 | `train: 0:1500` |
+| 时间步 | 视频 | 大小 | 总时长 | 训练划分 |
+| ---: | --- | ---: | ---: | --- |
+| **789,612** | 4,500 个 · AV1 / H.264 | ≈45.47 GB | ≈8.77 小时 | `train: 0:1500` |
 
-数据来源为 [RoboDojo](https://huggingface.co/datasets/RoboDojo-Benchmark/RoboDojo) 的 **1,393 条轨迹**和 **107 条补充轨迹**。源集合与轨迹编号记录在 `provenance/episodes.parquet` 中。
+[**浏览文件 →**](https://huggingface.co/datasets/wzlllong/value-model-benchmark-training-data/tree/main/data/simulation/full-shot) · [下载与读取](#download-and-reading)
 
 ### 2.3 One-Shot
 
-**15 个任务 × 1 条 = 15 条轨迹，共 8,080 个时间步。** 采样率为 **25 FPS**，状态与动作均为 `float32[14]`；三路 **640 × 480** RGB 相机共 **45 个 AV1 视频**。划分为 `train: 0:15`，大小约 **0.50 GB**。
+**15 个任务 × 1 条 = 15 条轨迹。** 与 Full-Shot 使用同一组任务定义。
 
-源任务名 `sweep_objects` 对应 Full-Shot 的 `sweep_blocks`，指令为用扫帚和簸箕扫积木。[任务—轨迹—行范围映射](https://huggingface.co/datasets/wzlllong/value-model-benchmark-training-data/blob/main/data/simulation/one-shot/provenance/episode_mapping.json)包含源任务名、episode 索引和数据行边界。
+| 时间步 | 视频 | 大小 | 总时长 | 训练划分 |
+| ---: | --- | ---: | ---: | --- |
+| **8,080** | 45 个 · AV1 | ≈0.50 GB | ≈5.39 分钟 | `train: 0:15` |
 
-读取前缀为 `data/simulation/one-shot`，规模为 **15 条轨迹、8,080 个时间步**。
+[**浏览文件 →**](https://huggingface.co/datasets/wzlllong/value-model-benchmark-training-data/tree/main/data/simulation/one-shot) · [任务—轨迹—行范围映射](https://huggingface.co/datasets/wzlllong/value-model-benchmark-training-data/blob/main/data/simulation/one-shot/provenance/episode_mapping.json) · [下载与读取](#download-and-reading)
 
-## 3. 真实世界
+## 3. 真实世界 <a name="real-world"></a>
 
 ### 3.1 总览
 
-真实世界数据由 **G2A 机器人**采集，包含 **20 个操作任务**，覆盖物体放置与分类、家居整理、抽屉操作、重复按压、厨房操作及包装。
-
-| 设置 | 每任务 | 总轨迹数 | 数据 |
-| --- | ---: | ---: | --- |
-| [**Full-Shot**](#32-full-shot) | 100 | **2,000** | [文件](https://huggingface.co/datasets/wzlllong/value-model-benchmark-training-data/tree/main/data/real-world/full-shot) |
-| [**One-Shot**](#33-one-shot) | 1 | **20** | [文件](https://huggingface.co/datasets/wzlllong/value-model-benchmark-training-data/tree/main/data/real-world/one-shot) |
-
-<details>
-<summary><strong>任务清单：20 个任务及简要说明</strong></summary>
-
-索引对���训练数据中的 `task_index`。名称与简述采用 RoboValue 论文附录 D 的任务定义；`meta/tasks.parquet` 中的原始任务字符串保持不变。
-
-| 索引 | 任务与简述 |
-| --- | --- |
-| 0 | **用盘子转移积木**<br>按红、绿、蓝顺序装盘，再倒入盒中。 |
-| 1 | **输入键盘数字序列**<br>依次按 2-8-2-6-8，每次松开并返回等待位置。 |
-| 2 | **摆放餐具**<br>从箱中取出盘、碗、叉、刀，按目标布局摆放。 |
-| 3 | **整理调料罐**<br>将三个调料罐放上双层架。 |
-| 4 | **将塑料盘放入盘架**<br>将三个塑料盘插入盘架。 |
-| 5 | **按按钮五次**<br>每次完全松开，并返回等待位置。 |
-| 6 | **将锅放到电磁炉上**<br>将锅放上电磁炉，收纳锅铲并盖上锅盖。 |
-| 7 | **整理桌面工作区**<br>收纳笔和橡皮、关闭抽屉，关灯后拔下电源。 |
-| 8 | **整理书本和笔**<br>对齐第一本书，将笔放入另一本书后合上并叠放。 |
-| 9 | **摆放家庭茶具**<br>将茶杯、茶壶及壶盖、茶罐和茶夹摆到目标位置。 |
-| 10 | **按顺序操作抽屉**<br>自下向上依次打开，再自上向下关闭。 |
-| 11 | **按顺序分类积木**<br>依次将红、黄、蓝积木放到指定抽屉柜上。 |
-| 12 | **装满并摆放笔筒**<br>将四支笔逐一插入笔筒，再将笔筒直立放下。 |
-| 13 | **整理化妆品收纳盒**<br>收纳口红与眉笔并关闭抽屉，再摆放精华和卸妆油。 |
-| 14 | **击鼓五次**<br>拿起鼓槌连续敲鼓五次，再放回鼓槌。 |
-| 15 | **重复按压洗洁精**<br>在出液口下拿住抹布，连续按压五次。 |
-| 16 | **反复给肉串撒料**<br>用调料瓶连续向肉串撒料五次。 |
-| 17 | **打包物品寄送**<br>先装鼠标，再装有线耳机，最后盖上气泡膜。 |
-| 18 | **插花并摆放藤条香薰**<br>连续插入三朵花，摆放香薰瓶并插入藤条。 |
-| 19 | **准备早餐**<br>按指定位置放入包子，叠加蒸笼层并盖上盖子。 |
-
-</details>
-
-### 3.2 Full-Shot
-
-**20 个任务 × 100 条 = 2,000 条轨迹，共 1,530,987 个时间步。**
-
-[下载与读取](#5-下载与读取) · [浏览文件](https://huggingface.co/datasets/wzlllong/value-model-benchmark-training-data/tree/main/data/real-world/full-shot) · [字段定义](https://huggingface.co/datasets/wzlllong/value-model-benchmark-training-data/blob/main/data/real-world/full-shot/meta/info.json)
-
-**关键统计**
-
-| 项目 | 数值 |
-| --- | --- |
-| 规模 | **20 个任务 × 100 条 = 2,000 条轨迹** |
-| 时间步与采样率 | **1,530,987** 个时间步，**30 FPS** |
-| 轨迹时长 | 总计约 **14.18 小时**；平均每条约 **25.52 秒** |
-| 相机 | **3 路 RGB：**头部、左腕、右腕 |
-| 分辨率 | 头部 **640 × 400**；每个腕部 **640 × 528** |
-| 存储 | 约 **23.23 GB**；Parquet + HEVC 视频 |
-| 划分 | `train: 0:2000` |
-
-时长与时间步按每条轨迹统计一次。
+真实世界数据由 **AgiBot Genie02** 机器人采集，覆盖 **20 个操作任务**，包括物体放置与分类、家居整理、抽屉操作、重复按压、厨房操作及包装，采样率为 **30 FPS**。
 
 **一条轨迹包含什么？**
 
-一条 **episode** 是一条完整的任务示例，包含任务与轨迹标识、逐时间步的机器人观测和动作，以及三路相机视频。轨迹元数据记录长度、对应的数据文件、数据行边界及各相机视频片段位置。
+Full-Shot 与 One-Shot 中，每条 **episode** 都是一条完整的任务示例，包含任务指令、逐时间步机器人观测与动作、轨迹标识和时间戳，以及头部、左腕和右腕三路 RGB 视频。轨迹元数据记录长度、数据行范围及各相机的视频片段位置。
 
-| 信号 | 表示 |
+| 内容 | 表示 |
 | --- | --- |
-| `action` | **52 维 float32：**同时包含**关节目标与末端目标位姿**，另含夹爪、头部、腰部和底盘控制字段 |
-| `observation.state` | **290 维 float32：**关节与末端状态，包括位置、朝向、速度、力矩／力及辅助字段 |
-| RGB 观测 | 头部、左腕、右腕图像；视频字段的图像元数据使用 CHW 顺序 |
-| 标识与时间 | `episode_index`、`task_index`、`frame_index`，以及单位为秒的逐帧 `timestamp` |
+| 机器人状态 | `observation.state`：`float32[290]`，包含关节与末端状态及辅助字段 |
+| 机器人动作 | `action`：`float32[52]`，包含关节目标、末端目标位姿及其他控制字段 |
+| RGB 观测 | 头部 **640 × 400**；左腕与右腕各 **640 × 528** |
+| 指令与标识 | 任务指令、`task_index`、`episode_index`、`frame_index` |
+| 时间 | `timestamp`，单位为秒；**30 FPS** |
 
-**末端动作目标由关节 action 通过正运动学（FK）计算。** 末端位置采用 `base_link` 坐标系，单位为 m，朝向为 XYZW 四元数。夹爪字段采用原始开度或行程值。52 维动作向量包含末端位姿、关节目标及其他控制字段，具体索引见下��。
+相机字段为 `observation.images.cam_high`、`observation.images.cam_left_wrist` 和 `observation.images.cam_right_wrist`。读取后的图像采用 **CHW** 顺序。
 
 <details>
 <summary><strong>动作参考：52 维完整索引</strong></summary>
 
+**末端动作目标由关节 action 通过正运动学（FK）计算。** 末端位置采用 `base_link` 坐标系，单位为 m，朝向为 XYZW 四元数。夹爪字段采用原始开度或行程值。52 维动作向量包含末端位姿、关节目标及其他控制字段，具体索引见下表。
+
 以下为 Python 切片，右端不包含在内。
 
 | 切片 | 含义 |
@@ -201,88 +171,136 @@
 
 </details>
 
+[Full-Shot 字段定义](https://huggingface.co/datasets/wzlllong/value-model-benchmark-training-data/blob/main/data/real-world/full-shot/meta/info.json) · [One-Shot 字段定义](https://huggingface.co/datasets/wzlllong/value-model-benchmark-training-data/blob/main/data/real-world/one-shot/meta/info.json)
+
+<details>
+<summary><strong>任务清单：20 个任务及简要说明</strong></summary>
+
+索引对应训练数据中的 `task_index`。名称与简述采用 RoboValue 论文附录 D 的任务定义；`meta/tasks.parquet` 中的原始任务字符串保持不变。
+
+| 索引 | 任务与简述 |
+| --- | --- |
+| 0 | **用盘子转移积木**<br>按红、绿、蓝顺序装盘，再倒入盒中。 |
+| 1 | **输入键盘数字序列**<br>依次按 2-8-2-6-8，每次松开并返回等待位置。 |
+| 2 | **摆放餐具**<br>从箱中取出盘、碗、叉、刀，按目标布局摆放。 |
+| 3 | **整理调料罐**<br>将三个调料罐放上双层架。 |
+| 4 | **将塑料盘放入盘架**<br>将三个塑料盘插入盘架。 |
+| 5 | **按按钮五次**<br>每次完全松开，并返回等待位置。 |
+| 6 | **将锅放到电磁炉上**<br>将锅放上电磁炉，收纳锅铲并盖上锅盖。 |
+| 7 | **整理桌面工作区**<br>收纳笔和橡皮、关闭抽屉，关灯后拔下电源。 |
+| 8 | **整理书本和笔**<br>对齐第一本书，将笔放入另一本书后合上并叠放。 |
+| 9 | **摆放家庭茶具**<br>将茶杯、茶壶及壶盖、茶罐和茶夹摆到目标位置。 |
+| 10 | **按顺序操作抽屉**<br>自下向上依次打开，再自上向下关闭。 |
+| 11 | **按顺序分类积木**<br>依次将红、黄、蓝积木放到指定抽屉柜上。 |
+| 12 | **装满并摆放笔筒**<br>将四支笔逐一插入笔筒，再将笔筒直立放下。 |
+| 13 | **整理化妆品收纳盒**<br>收纳口红与眉笔并关闭抽屉，再摆放精华和卸妆油。 |
+| 14 | **击鼓五次**<br>拿起鼓槌连续敲鼓五次，再放回鼓槌。 |
+| 15 | **重复按压洗洁精**<br>在出液口下拿住抹布，连续按压五次。 |
+| 16 | **反复给肉串撒料**<br>用调料瓶连续向肉串撒料五次。 |
+| 17 | **打包物品寄送**<br>先装鼠标，再装有线耳机，最后盖上气泡膜。 |
+| 18 | **插花并摆放藤条香薰**<br>连续插入三朵花，摆放香薰瓶并插入藤条。 |
+| 19 | **准备早餐**<br>按指定位置放入包子，叠加蒸笼层并盖上盖子。 |
+
+</details>
+
+### 3.2 Full-Shot
+
+**20 个任务 × 100 条 = 2,000 条轨迹。**
+
+| 时间步 | 视频 | 大小 | 总时长 | 训练划分 |
+| ---: | --- | ---: | ---: | --- |
+| **1,530,987** | 6,000 个 · HEVC | ≈23.23 GB | ≈14.18 小时 | `train: 0:2000` |
+
+[**浏览文件 →**](https://huggingface.co/datasets/wzlllong/value-model-benchmark-training-data/tree/main/data/real-world/full-shot) · [下载与读取](#download-and-reading)
+
 ### 3.3 One-Shot
 
-**20 个任务 × 1 条 = 20 条轨迹，共 15,667 个时间步。** 采样率为 **30 FPS**，状态为 `float32[290]`，动作为 `float32[52]`；三路 RGB 相机共 **60 个 HEVC（H.265）视频**。划分为 `train: 0:20`，大小约 **0.23 GB**。
+**20 个任务 × 1 条 = 20 条轨迹。** 与 Full-Shot 使用同一组任务定义。
 
-[任务—轨迹—行范围映射](https://huggingface.co/datasets/wzlllong/value-model-benchmark-training-data/blob/main/data/real-world/one-shot/provenance/episode_mapping.json)包含源任务标识、episode 索引和数据行边界。相机字段为 `cam_left_wrist`、`cam_right_wrist` 和 `cam_high`，对应左腕、右腕与头部视角。
+| 时间步 | 视频 | 大小 | 总时长 | 训练划分 |
+| ---: | --- | ---: | ---: | --- |
+| **15,667** | 60 个 · HEVC | ≈0.23 GB | ≈8.70 分钟 | `train: 0:20` |
 
-读取前缀为 `data/real-world/one-shot`，规模为 **20 条轨迹、15,667 个时间步**。
+[**浏览文件 →**](https://huggingface.co/datasets/wzlllong/value-model-benchmark-training-data/tree/main/data/real-world/one-shot) · [任务—轨迹—行范围映射](https://huggingface.co/datasets/wzlllong/value-model-benchmark-training-data/blob/main/data/real-world/one-shot/provenance/episode_mapping.json) · [下载与读取](#download-and-reading)
 
-## 4. 仓库目录
+## 4. 仓库目录 <a name="repository-layout"></a>
 
 ```text
 README.md                         英文数据集���片
 README_zh.md                      中文数据集卡片
-subset_catalog.json               机器可读子集索引
+subset_catalog.json               子集索引
 data/
 ├── simulation/
-│   ├── full-shot/                LeRobot 数据根目录
-│   └── one-shot/                 LeRobot 数据根目录
+│   ├── full-shot/                1,500 条 · 15 个任务
+│   └── one-shot/                 15 条 · 每任务 1 条
 └── real-world/
-    ├── full-shot/                LeRobot 数据根目录
-    └── one-shot/                 LeRobot 数据根目录
+    ├── full-shot/                2,000 条 · 20 个任务
+    └── one-shot/                 20 条 · 每任务 1 条
 ```
 
-每个子集目录分别作为 LeRobot 的 `root`；四个子集均采用上表所示的目录结构。
+每个子集目录都是独立的 **LeRobot `root`**，内部包含逐时间步 Parquet、三路原始相机视频及元数据。
 
 <details>
-<summary><strong>格式参考：任一已发布子集内部的文件</strong></summary>
+<summary><strong>子集文件结构与元数据</strong></summary>
 
 | 文件或目录 | 用途 |
 | --- | --- |
 | `data/chunk-*/file-*.parquet` | 逐时间步状态、动作、标识与时间戳 |
 | `videos/<camera>/chunk-*/file-*.mp4` | RGB 相机视频流 |
-| `meta/info.json` | v3.0 版本、字段、维度、采样率与路径模板 |
-| `meta/tasks.parquet` | 原始任务字符串及其索引 |
-| `meta/episodes/chunk-*/file-*.parquet` | 轨迹长度、任务、数据边界、视频偏移与逐轨迹统计 |
+| `meta/info.json` | 格式版本、字段、维度、采样率与路径模板 |
+| `meta/tasks.parquet` | 任务指令及索引 |
+| `meta/episodes/chunk-*/file-*.parquet` | 轨迹长度、任务、数据行边界、视频偏移与逐轨迹统计 |
 | `meta/stats.json` | 数据集级别的字段统计 |
 | `provenance/` | 轨迹来源与任务对应关系 |
 | `checksums.sha256` | 文件完整性清单 |
 
-相机字段为 `observation.images.cam_high`、`observation.images.cam_left_wrist` 和 `observation.images.cam_right_wrist`。轨迹数据行边界为 `dataset_from_index` 与 `dataset_to_index`，后者不包含在内；各相机视频的时间段分别记录。
+数据行范围为 `[dataset_from_index, dataset_to_index)`；各相机的视频时间段分别记录。时间步与时长按每条轨迹统计一次。
 
-[LeRobot Dataset v3.0 官方格式说明](https://huggingface.co/docs/lerobot/lerobot-dataset-v3) · [子集索引](https://huggingface.co/datasets/wzlllong/value-model-benchmark-training-data/blob/main/subset_catalog.json)
+[LeRobot v3.0 官方格式说明](https://huggingface.co/docs/lerobot/lerobot-dataset-v3) · [机器可读子集索引](https://huggingface.co/datasets/wzlllong/value-model-benchmark-training-data/blob/main/subset_catalog.json)
 
 </details>
 
-## 5. 下载与读取
+## 5. 下载与读取 <a name="download-and-reading"></a>
 
 ### 5.1 选择子集
 
-以下四个前缀分别是独立的 LeRobot 根目录。存储大小按十进制 GB 计算，包含原始视频。
+只需选择一个 `prefix`；下载范围限定在该子集内。大小采用十进制 GB。
 
-| 子集 | `prefix` | 轨迹 | 数据行 | 视频 | FPS | 状态 / 动作维度 | 大小 |
-| --- | --- | ---: | ---: | ---: | ---: | --- | ---: |
-| 仿真 Full-Shot | `data/simulation/full-shot` | 1,500 | 789,612 | 4,500 | 25 | 14 / 14 | ≈45.47 GB |
-| 仿真 One-Shot | `data/simulation/one-shot` | 15 | 8,080 | 45 | 25 | 14 / 14 | ≈0.50 GB |
-| 真机 Full-Shot | `data/real-world/full-shot` | 2,000 | 1,530,987 | 6,000 | 30 | 290 / 52 | ≈23.23 GB |
-| 真机 One-Shot | `data/real-world/one-shot` | 20 | 15,667 | 60 | 30 | 290 / 52 | ≈0.23 GB |
+| 子集 | `prefix` | 轨迹 / 时间步 | 大小 |
+| --- | --- | ---: | ---: |
+| 仿真 Full-Shot | `data/simulation/full-shot` | 1,500 / 789,612 | ≈45.47 GB |
+| 仿真 One-Shot | `data/simulation/one-shot` | 15 / 8,080 | ≈0.50 GB |
+| 真机 Full-Shot | `data/real-world/full-shot` | 2,000 / 1,530,987 | ≈23.23 GB |
+| 真机 One-Shot | `data/real-world/one-shot` | 20 / 15,667 | ≈0.23 GB |
 
 ### 5.2 下载并读取
 
-**示例默认读取仿真 Full-Shot。** 若希望先做较小规模的下载，可将 `prefix` 改为任一 One-Shot 子集。公开下载无需 Token。
+以下示例读取较小的 **仿真 One-Shot**，返回状态、动作及三路相机帧。切换子集时，将 `prefix` 改为上表中的对应路径。公开下载无需 Token。
 
-读取环境采用 **Python 3.10+**、**LeRobot 0.4.2** 和 **PyAV 15.1.0**。数据格式版本为 **LeRobot Dataset v3.0**。
+使用 **Python 3.10+**、**LeRobot 0.4.2** 与 **PyAV 15.1.0**：
 
 ```bash
-pip install "lerobot==0.4.2" "av==15.1.0"
+python -m pip install "lerobot==0.4.2" "av==15.1.0"
 ```
 
 ```python
+import os
 from pathlib import Path
 
 from huggingface_hub import snapshot_download
 from lerobot.datasets.lerobot_dataset import LeRobotDataset
 
 repo_id = "wzlllong/value-model-benchmark-training-data"
-prefix = "data/simulation/full-shot"
+prefix = "data/simulation/one-shot"
 revision = "3a81d8f75444d57ecae1d604de0d604130ec60cd"
+local_dir = Path(os.environ.get("ROBOVALUE_LOCAL_DIR", "./robovalue-data"))
+
 snapshot = snapshot_download(
     repo_id=repo_id,
     repo_type="dataset",
     revision=revision,
     allow_patterns=[f"{prefix}/**"],
+    local_dir=local_dir,
     token=False,
 )
 dataset = LeRobotDataset(
@@ -292,16 +310,31 @@ dataset = LeRobotDataset(
     video_backend="pyav",
 )
 sample = dataset[0]
-print(dataset.num_episodes, dataset.num_frames)
-print(sample["action"].shape, sample["observation.state"].shape)
+print("episodes / frames:", dataset.num_episodes, dataset.num_frames)
+print("action / state:", tuple(sample["action"].shape),
+      tuple(sample["observation.state"].shape))
+for camera in dataset.meta.camera_keys:
+    print(camera, tuple(sample[camera].shape))
 ```
 
-该示例返回 **1,500 条轨迹、789,612 个时间步**，`action` 与 `observation.state` 的维度均为 **[14]**。`dataset[0]` 返回第一个时间步的状态、动作和对应相机帧。
+输出：
 
-下载和读取使用同一个固定版本，只下载所选前缀下的文件。四个子集的读取前缀、轨迹数和数据行数见第 5.1 节。
+```text
+episodes / frames: 15 8080
+action / state: (14,) (14,)
+observation.images.cam_high (3, 480, 640)
+observation.images.cam_left_wrist (3, 480, 640)
+observation.images.cam_right_wrist (3, 480, 640)
+```
+
+文件保存在 `./robovalue-data`。设置环境变量 `ROBOVALUE_LOCAL_DIR` 可指定已有下载目录，复用本地文件。下载与读取采用同一个固定数据版本。
 
 ### 5.3 完整性检查
 
-每个根目录均包含 `checksums.sha256`。下载后可在子集目录内运行 `shasum -a 256 -c checksums.sha256` 检查完整性。源轨迹与任务映射见 `provenance/`，字段定义和统计量分别见 `meta/info.json` 与 `meta/stats.json`。
+每个子集包含 `checksums.sha256`，可在对应子集目录中运行：
+
+```bash
+shasum -a 256 -c checksums.sha256
+```
 
-**发布授权：**本数据集经授权公开发布。RoboDojo 来源部分遵循 Apache-2.0 许可。
+**发布授权：**本数据集全部数据均经授权后公开发布。

subset_catalog.json

@@ -14,7 +14,8 @@
       "subsets": [
         "simulation-full-shot-data",
         "simulation-one-shot-data"
-      ]
+      ],
+      "platform": "ARX"
     },
     {
       "domain": "real-world",
@@ -23,7 +24,8 @@
       "subsets": [
         "real-world-full-shot-data",
         "real-world-one-shot-data"
-      ]
+      ],
+      "platform": "AgiBot Genie02"
     }
   ],
   "available_subsets": [
@@ -57,7 +59,27 @@
         "av1",
         "h264"
       ],
-      "videos": 4500
+      "videos": 4500,
+      "platform": "ARX",
+      "source_description_en": "Based on open-source RoboDojo data, reviewed by our team with selected demonstrations recollected.",
+      "source_description_zh": "基于 RoboDojo 开源数据，经团队审核后，对部分示例重新采集。",
+      "display_task_names": [
+        "arrange_largest_number",
+        "fill_pen_holder",
+        "fold_clothes",
+        "hang_mugs",
+        "insert_tubes",
+        "make_toast",
+        "organize_table",
+        "pack_objects_into_box",
+        "play_stacking_toy",
+        "press_by_number",
+        "put_bottles_into_dustbin",
+        "stack_blocks",
+        "store_laptop_and_headphones",
+        "swap_blocks",
+        "sweep_blocks"
+      ]
     },
     {
       "subset_id": "simulation-one-shot-data",
@@ -84,10 +106,27 @@
         "av1"
       ],
       "episode_mapping": "data/simulation/one-shot/provenance/episode_mapping.json",
-      "task_name_correspondence": {
-        "sweep_objects": "sweep_blocks"
-      },
-      "videos": 45
+      "videos": 45,
+      "platform": "ARX",
+      "source_description_en": "Based on open-source RoboDojo data, reviewed by our team with selected demonstrations recollected.",
+      "source_description_zh": "基于 RoboDojo 开源数据，经团队审核后，对部分示例重新采集。",
+      "display_task_names": [
+        "arrange_largest_number",
+        "fill_pen_holder",
+        "fold_clothes",
+        "hang_mugs",
+        "insert_tubes",
+        "make_toast",
+        "organize_table",
+        "pack_objects_into_box",
+        "play_stacking_toy",
+        "press_by_number",
+        "put_bottles_into_dustbin",
+        "stack_blocks",
+        "store_laptop_and_headphones",
+        "swap_blocks",
+        "sweep_blocks"
+      ]
     },
     {
       "subset_id": "real-world-full-shot-data",
@@ -115,7 +154,8 @@
         "end-effector-target-poses"
       ],
       "end_effector_target_source": "forward kinematics from joint actions",
-      "videos": 6000
+      "videos": 6000,
+      "platform": "AgiBot Genie02"
     },
     {
       "subset_id": "real-world-one-shot-data",
@@ -143,9 +183,11 @@
       ],
       "episode_mapping": "data/real-world/one-shot/provenance/episode_mapping.json",
       "source_episode_selection": "original episode 000 per source task",
-      "videos": 60
+      "videos": 60,
+      "platform": "AgiBot Genie02"
     }
   ],
   "planned_subsets": [],
-  "data_revision": "3a81d8f75444d57ecae1d604de0d604130ec60cd"
+  "data_revision": "3a81d8f75444d57ecae1d604de0d604130ec60cd",
+  "publication_authorization": "All data were authorized before public release."
 }