# 🟦 Simulation One-Shot Training Data / 仿真 One-Shot 训练集

[Dataset Card](https://huggingface.co/datasets/wzlllong/value-model-benchmark-training-data#simulation) · [简体中文](https://huggingface.co/datasets/wzlllong/value-model-benchmark-training-data/blob/main/README_zh.md#simulation) · [Download & Read](https://huggingface.co/datasets/wzlllong/value-model-benchmark-training-data#download-and-reading)

**15 tasks × 1 demonstration per task = 15 episodes.** All episodes belong to the `train` split (`0:15`).

Based on open-source [RoboDojo](https://huggingface.co/datasets/RoboDojo-Benchmark/RoboDojo) data reviewed by our team, with selected demonstrations recorded again.

| Content | Specification |
| --- | --- |
| Platform | ARX |
| Dataset format | LeRobot Dataset v3.0 |
| Synchronized time steps | 8,080 |
| Sampling rate | 25 FPS |
| Robot state | `observation.state`: `float32[14]` |
| Robot action | `action`: `float32[14]` |
| RGB cameras | High, left wrist, right wrist: 640 × 480 each |
| Camera videos | 45 · AV1 |

This directory is an independent LeRobot dataset root. Set `root` to `data/simulation/one-shot` inside the downloaded snapshot. The [download and training examples](https://huggingface.co/datasets/wzlllong/value-model-benchmark-training-data#download-and-reading) use LeRobot **0.4.2** and PyAV **15.1.0** and return states, actions, task instructions and three camera observations as PyTorch batches. Value or progress supervision is constructed by the training method.

[Feature definitions](https://huggingface.co/datasets/wzlllong/value-model-benchmark-training-data/blob/main/data/simulation/one-shot/meta/info.json) · [Shared trajectory contents and task descriptions](https://huggingface.co/datasets/wzlllong/value-model-benchmark-training-data#simulation)

<details>
<summary><strong>Task–episode–row mapping</strong></summary>

| Task index | Source task key | Episodes | Episode range / index | Data-row range [start, end) |
| ---: | --- | ---: | --- | --- |
| 0 | `arrange_largest_number` | 1 | 0 | [0, 579) |
| 1 | `fill_pen_holder` | 1 | 1 | [579, 1513) |
| 2 | `fold_clothes` | 1 | 2 | [1513, 1814) |
| 3 | `hang_mugs` | 1 | 3 | [1814, 2327) |
| 4 | `insert_tubes` | 1 | 4 | [2327, 2647) |
| 5 | `make_toast` | 1 | 5 | [2647, 3231) |
| 6 | `organize_table` | 1 | 6 | [3231, 4070) |
| 7 | `pack_objects_into_box` | 1 | 7 | [4070, 4484) |
| 8 | `play_stacking_toy` | 1 | 8 | [4484, 5372) |
| 9 | `press_by_number` | 1 | 9 | [5372, 5750) |
| 10 | `put_bottles_into_dustbin` | 1 | 10 | [5750, 6113) |
| 11 | `stack_blocks` | 1 | 11 | [6113, 6381) |
| 12 | `store_laptop_and_headphones` | 1 | 12 | [6381, 6932) |
| 13 | `swap_blocks` | 1 | 13 | [6932, 7377) |
| 14 | `sweep_blocks` | 1 | 14 | [7377, 8080) |

</details>

Task indices match `meta/tasks.parquet`. Source episode identifiers and task correspondence are recorded in `provenance/`. `checksums.sha256` provides file integrity hashes.

**Publication authorization:** All data were authorized before public release.

## ☰ Subtask annotations

[Supplementary subtask annotations](https://huggingface.co/datasets/wzlllong/value-model-benchmark-training-data/blob/main/data/simulation/one-shot/annotations/subtasks/README.md) provide **58 paper-defined segments and 8,080 frame labels**, one active subtask per time step. Labels use `episode_index` and `frame_index`, with `global_index` matching the LeRobot `index`. Initial waiting is part of the first subtask, and the final subtask extends to the episode end. Segments use `[start_frame, end_frame)`; the included last frame is `end_frame - 1`.

Read these files explicitly using PyArrow or `load_subtasks.py`; they are separate from the LeRobot samples and benchmark test-set labels. See the directory README for the fixed-revision download example and boundary adjustment mask. [`dataset_labeler`-compatible JSON](https://huggingface.co/datasets/wzlllong/value-model-benchmark-training-data/tree/7fd21cb2519b4d59cbe9fe8b23cfdf477e3ad145/data/simulation/one-shot/annotations/subtasks/dataset_labeler) uses the same subtasks and boundaries as the frame labels.

## 中文

**15 个任务 × 每任务 1 条 = 15 条轨迹**，共 **8,080 个同步时间步**，全部属于 `train` 划分（`0:15`）。

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。

平台为 **ARX**，采用 **LeRobot Dataset v3.0**，采样率为 **25 FPS**。状态为 `float32[14]`，动作为 `float32[14]`；三路 RGB 相机为高位、左腕、右腕：均为 640 × 480，共 **45 个原编码视频（AV1）**。

下载后，将 LeRobot 的 `root` 设置为快照中的 `data/simulation/one-shot`。主卡片的[下载与训练示例](https://huggingface.co/datasets/wzlllong/value-model-benchmark-training-data/blob/main/README_zh.md#download-and-reading)提供状态、动作、任务指令与三路相机观测的 PyTorch batch；value 或 progress 监督由训练方法构造。

折叠表列出任务索引、轨迹分配及数据行范围；原始来源标识与对应关系保留在 `provenance/`。

**发布授权：**全部数据均经授权后公开发布。

## ☰ 子任务标注

[子任务标注](https://huggingface.co/datasets/wzlllong/value-model-benchmark-training-data/blob/main/data/simulation/one-shot/annotations/subtasks/README.md)提供 **58 个论文子任务区间、8,080 条逐帧标签**，每个时间步一个当前子任务。使用 `episode_index` 与 `frame_index` 关联，并核对 `global_index` 与 LeRobot 的 `index` 一致。起始等待归入第一个子任务，最后一个子任务持续至轨迹末帧；区间为 `[start_frame, end_frame)`，包含式末帧为 `end_frame - 1`。

通过 PyArrow 或 `load_subtasks.py` 显式读取。标注独立于 LeRobot sample，与评测集标签分开；目录 README 提供固定版本的下载示例及边界调整标记说明。 另提供 [`dataset_labeler` 兼容 JSON](https://huggingface.co/datasets/wzlllong/value-model-benchmark-training-data/tree/7fd21cb2519b4d59cbe9fe8b23cfdf477e3ad145/data/simulation/one-shot/annotations/subtasks/dataset_labeler)，与逐帧标签使用相同的子任务和边界。