---
title: RoboValue-Dataset
canonical_url: "https://www.modelscope.ai/datasets/RoboValue/RoboValue-Dataset"
md_url: "https://www.modelscope.ai/datasets/RoboValue/RoboValue-Dataset.md"
repository: RoboValue/RoboValue-Dataset
last_updated: 2026-11-12
license: other
downloads: 56
stars: 0
---

# RoboValue-Dataset

> RoboValue-Dataset - An open-source dataset by RoboValue on ModelScope. 简体中文 · Paper · Code · Project

RoboValue/RoboValue-Dataset is a dataset on ModelScope. licensed under other.

- **Repository**: RoboValue/RoboValue-Dataset
- **License**: other
- **Downloads**: 56
- **Stars**: 0
- **Last updated**: 2026-11-12

Source: https://www.modelscope.ai/datasets/RoboValue/RoboValue-Dataset

---

# RoboValue Dataset

[简体中文](README_zh.md) · [Paper](https://arxiv.org/abs/2610.26663) · [Code](https://github.com/RoboValue-Benchmark/RoboValue) · [Project](https://robovalue-benchmark.github.io/)

<img src="assets/robovalue-logo.png" alt="RoboValue" width="320" />

RoboValue provides expert demonstrations and small-scale diagnostic evaluation data for robotic value models, covering **15 simulation training tasks and 20 real-world training tasks** in **LeRobot Dataset v3.0** format. The four training subsets contain 3,535 episodes; the two evaluation subsets contain 60 trajectories.

This ModelScope repository is `gechang127/RoboValue-Dataset`. Videos, Parquet files, annotations and native LeRobot metadata retain their upstream content. Documentation, presentation assets, navigation and integrity manifests are adapted for this repository.

<a id="overview"></a>

## Dataset overview

Each subset can be downloaded independently. Sizes use decimal GB. Full-Shot contains 100 expert demonstrations per task; One-Shot contains one per task.

| Subset path | Tasks | Episodes | Synchronized time steps | Size |
| --- | ---: | ---: | ---: | ---: |
| [training-data/simulation/full-shot](training-data/simulation/full-shot/README.md) | 15 | 1,500 | 789,612 | 45.468 GB |
| [training-data/simulation/one-shot](training-data/simulation/one-shot/README.md) | 15 | 15 | 8,080 | 0.500 GB |
| [training-data/real-world/full-shot](training-data/real-world/full-shot/README.md) | 20 | 2,000 | 1,534,321 | 23.126 GB |
| [training-data/real-world/one-shot](training-data/real-world/one-shot/README.md) | 20 | 20 | 16,362 | 0.216 GB |
| [evaluation/simulation/test-30](evaluation/simulation/test-30/README.md) | 4 | 30 | 17,149 | 0.595 GB |
| [evaluation/real-world/test-30](evaluation/real-world/test-30/README.md) | 4 | 30 | 27,498 | 0.382 GB |

<a id="simulation"></a>

Simulation training uses ARX at 25 FPS, with 14-dimensional states and actions.

<a id="real-world"></a>

Real-world training uses AgiBot Genie02 at 30 FPS, with 290-dimensional states and 52-dimensional actions. Both domains provide three synchronized camera views. Task lists, feature definitions and camera specifications are in the [data format and usage guide](docs/guide.md).

<a id="public-evaluation"></a>

## Evaluation scope

Each `test-30` subset contains four tasks, 30 trajectories and 90 camera videos for small-scale diagnostic evaluation. **These subsets do not constitute the full hidden leaderboard test set.** Model inputs are images and English instructions; evaluation annotations are stored separately. Do not feed success labels or metric memberships to the model. Task and episode indices are local to each subset.

<a id="metric-summary"></a>

[Metric definitions](docs/guide.md#metric-summary) · [Simulation coverage](evaluation/simulation/test-30/README.md#metric-coverage) · [Real-world coverage](evaluation/real-world/test-30/README.md#metric-coverage) · [Simulation trajectory index](docs/evaluation/simulation-trajectories.md) · [Real-world trajectory index](docs/evaluation/real-world-trajectories.md)

Simulation TRR/MEM terminal labels retain the source status of pending independent verification; see the simulation subset card. Availability of subtask definitions and temporal labels is documented in [Subtask annotations](docs/subtasks.md).

## Dataset preview

The ModelScope **Preview** tab shows one row per episode across six configurations: four training subsets (`train`) and two diagnostic subsets (`test`). The four small tables in `preview/` are derived from native episode/task metadata; evaluation previews reuse each subset's `episode_catalog.parquet`. These tables are browsing indexes. Load per-frame observations, actions and videos from the unchanged LeRobot subset roots using the guide below.

Training preview paths are repository-relative. Evaluation catalog video paths are relative to the selected evaluation subset. Episode indices are local to each subset. Evaluation labels are for scoring and must not be used as model inputs.

<a id="download-and-reading"></a>

## Download from ModelScope

This example downloads only the smaller Simulation One-Shot subset. Change `prefix` to select another subset; omit `allow_patterns` to download the entire repository. Private repositories require a token with read access, supplied through `MODELSCOPE_API_TOKEN`. Replace `master` with a ModelScope commit ID to pin a version.

```bash
python -m pip install "modelscope-hub==0.6.1"
```

```python
import os
from modelscope_hub import HubApi

prefix = "training-data/simulation/one-shot"
api = HubApi(token=os.environ.get("MODELSCOPE_API_TOKEN"))
snapshot = api.download_repo(
    repo_id="gechang127/RoboValue-Dataset",
    repo_type="dataset",
    revision="master",
    allow_patterns=[f"{prefix}/**"],
    local_dir="./robovalue-data",
)
print(snapshot)
```

Set LeRobot's `root` to the selected subset inside `./robovalue-data/`. Dependency versions, LeRobot loading, PyTorch batches and evaluation reading examples are maintained in the [usage guide](docs/guide.md#download-and-reading).

<a id="repository-layout"></a>

## Layout and integrity

```text
RoboValue-Dataset/
├── README.md / README_zh.md
├── LICENSE
├── subset_catalog.json
├── checksums.sha256
├── training-data/
│   ├── simulation/{full-shot,one-shot}/
│   └── real-world/{full-shot,one-shot}/
├── evaluation/
│   ├── simulation/test-30/
│   └── real-world/test-30/
├── docs/
│   ├── guide.md / guide_zh.md
│   ├── subtasks.md / subtasks_zh.md
│   └── evaluation/{simulation,real-world}-trajectories.md
├── preview/
├── assets/
└── provenance/
```

Every subset retains `data/`, `videos/`, `meta/`, `annotations/`, its README and `checksums.sha256`. Evaluation subsets also contain `episode_catalog.parquet`, with one row per trajectory. See [subset_catalog.json](subset_catalog.json) for the machine-readable index.

Run `sha256sum -c checksums.sha256` inside a downloaded subset, or at the repository root after a full download. Current manifests cover this ModelScope edition. Original manifests in `provenance/huggingface/` preserve the upstream integrity record.

## License

This dataset, including videos, trajectories, annotations, metadata and documentation, is released under the [Apache License 2.0](LICENSE). The dataset publisher has confirmed that this license applies to the full dataset.

## Source and preview

Source: [Hugging Face revision c746556](https://huggingface.co/datasets/wzlllong/value-model-benchmark-training-data/tree/c74655607eeb8dfab4e37bde409b42f42eb12e08). Migration and documentation cleanup are recorded in [provenance/migration.json](provenance/migration.json). Please cite the [RoboValue paper](https://arxiv.org/abs/2610.26663) when using the dataset.

[![Real-world task preview](assets/previews/real-world-task-00.jpg)](assets/previews/real-world-task-00.mp4)
