README.md

---
license: cc-by-nc-4.0
task_categories:
  - visual-question-answering
  - robotics
  - image-to-text
language:
  - en
pretty_name: UnoBench
tags:
  - vision-language
  - robotic-grasping
  - obstruction-reasoning
  - cluttered-scenes
  - set-of-mark
  - benchmark
---

# UnoBench

**UnoBench** is a benchmark for target-centric obstruction reasoning in robotic grasping under cluttered scenes. Given a target object, a method must identify the objects that block or constrain access to that target before grasping.

UnoBench is built upon [MetaGraspNetV2](https://github.com/maximiliangilles/MetaGraspNet) and extends the initial idea of [FreeGraspData](https://huggingface.co/datasets/FBK-TeV/FreeGraspData).

## Resources

| Resource | Link | Description |
| --- | --- | --- |
| UnoGrasp code | [GitHub main branch](https://github.com/tev-fbk/UnoGrasp/tree/main) | Method code, checkpoints, inference, and evaluation. |
| Challenge starter kit | [GitHub challenge branch](https://github.com/tev-fbk/unobenchchallenge/tree/master) | Minimal examples and local evaluators for the UnoBench Challenge. |
| Project page | [tev-fbk.github.io/UnoGrasp](https://tev-fbk.github.io/UnoGrasp/) | Paper, video, and release links. |

## Dataset Overview

<img src="benchmark.jpg" alt="UnoBench teaser" width="1500">

UnoBench provides synthetic cluttered-scene data with RGB images, Set-of-Mark images, instance annotations, natural-language object descriptions, and obstruction metadata.

The benchmark supports two settings:

| Setting | Target input | Expected output | Main use case |
| --- | --- | --- | --- |
| NLP | Natural-language target description and RGB image | Obstructing objects, represented by points or grounded object IDs | Vision-language and language-conditioned methods. |
| SoM | Set-of-Mark image and target object ID | Obstructing object IDs | Object-centric, graph-based, or modular robotic reasoning methods. |

## Dataset Structure

```text
UnoBench/
`-- UnoBenchSyn/
    |-- images.zip
    |-- images_som.zip
    |-- annotations.zip
    |-- test_GT_small.json
    |-- test_nlp_small.jsonl
    |-- test_som_small.jsonl
    |-- challenge_only/
    |   |-- test_nlp.jsonl
    |   `-- test_som.jsonl
    `-- meta_data/
        |-- Synthetic_train.json
        |-- image_id_scene_view_id_mapping.json
        |-- name_for_all.json
        |-- annotations_meta.zip
        `-- occ_info/
            |-- obs_information.json
            `-- masks.zip
```

After extracting the main archives, the dataset also contains:

```text
UnoBenchSyn/
|-- images/        # RGB images
|-- images_som/    # Set-of-Mark images
`-- annotations/   # Instance masks used by NLP point evaluation
```

## File Description

### Main Archives

| File | Description |
| --- | --- |
| `images.zip` | RGB images of synthetic cluttered scenes. |
| `images_som.zip` | Set-of-Mark images with object IDs / visual prompts. |
| `annotations.zip` | Instance segmentation masks associated with each image. These masks map image points to object IDs. |

### Reproduction Files

These files are used by the [UnoGrasp code](https://github.com/tev-fbk/UnoGrasp/tree/main) for inference and evaluation on the released small split.

| File | Description |
| --- | --- |
| `test_GT_small.json` | Ground-truth obstruction annotations for the small test split. |
| `test_som_small.jsonl` | Evaluation samples for the SoM setting. |
| `test_nlp_small.jsonl` | Evaluation samples for the NLP setting. |

### Challenge Files

These files are used by the [UnoBench Challenge](https://unobenchchallenge.fbk.eu/).

| File | Description |
| --- | --- |
| `challenge_only/test_som.jsonl` | Challenge Track 1: Set-of-Mark reasoning. |
| `challenge_only/test_nlp.jsonl` | Challenge Track 2: natural-language reasoning. |

### Metadata

| File | Description |
| --- | --- |
| `meta_data/Synthetic_train.json` | Query object, target objects, occlusion paths, and difficulty level for each sample. |
| `meta_data/image_id_scene_view_id_mapping.json` | Mapping between image IDs, scene IDs, and view IDs. |
| `meta_data/name_for_all.json` | Human-annotated object descriptions. |
| `meta_data/annotations_meta.zip` | MetaGraspNetV2 annotations, including depth, semantic segmentation, instance segmentation, and occlusion masks. |
| `meta_data/occ_info/obs_information.json` | Pairwise obstruction/occlusion information, such as obstruction ratio, contact point, and obstruction degree. |
| `meta_data/occ_info/masks.zip` | Instance masks for obstruction pairs. |

## Download

Install the Hugging Face CLI if needed:

```bash
pip install -U huggingface_hub
```

Download the full dataset:

```bash
hf download FBK-TeV/UnoBench \
  --repo-type dataset \
  --local-dir ./UnoBench/UnoBenchSyn
```

Or download individual archives:

```bash
hf download FBK-TeV/UnoBench images.zip \
  --repo-type dataset \
  --local-dir ./UnoBench/UnoBenchSyn

hf download FBK-TeV/UnoBench images_som.zip \
  --repo-type dataset \
  --local-dir ./UnoBench/UnoBenchSyn

hf download FBK-TeV/UnoBench annotations.zip \
  --repo-type dataset \
  --local-dir ./UnoBench/UnoBenchSyn
```

## Extraction

After downloading, unzip the main archives:

```bash
cd UnoBench/UnoBenchSyn
unzip images.zip
unzip images_som.zip
unzip annotations.zip
```

## Evaluation Splits

| Split / file | Purpose |
| --- | --- |
| `test_som_small.jsonl` | SoM reproduction with the released UnoGrasp small checkpoint. |
| `test_nlp_small.jsonl` | NLP reproduction with the released UnoGrasp small checkpoint. |
| `test_GT_small.json` | Ground truth for local reproduction evaluation. |
| `challenge_only/test_som.jsonl` | Official challenge queries for the SoM track. |
| `challenge_only/test_nlp.jsonl` | Official challenge queries for the NLP track. |

The challenge test ground truth is reserved for official leaderboard evaluation.

## Metadata Format

Metadata files provide scene-level and object-level information, including:

```text
image_id
scene_id
view_id
query_object
target_object
occlusion_paths
difficulty
num_paths
k_min
som_only
```

The obstruction information is target-centric: for each target object, UnoBench describes the objects that obstruct it and the corresponding obstruction paths.

## Notes

UnoBench focuses on high-level obstruction reasoning before grasping, rather than low-level grasp pose execution or robot control. In this release, obstruction is operationalized mainly through occlusion relationships in cluttered scenes.

## Citation

If you use UnoBench in your research, please cite:

```bibtex
@inproceedings{jiao2026obstruction,
  title = {Obstruction Reasoning for Robotic Grasping},
  author = {Runyu Jiao and Matteo Bortolon and Francesco Giuliari and Alice Fasoli and Sergio Povoli and Guofeng Mei and Yiming Wang and Fabio Poiesi},
  booktitle = {Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)},
  year = {2026}
}
```

## License

UnoBench is released under the **CC BY-NC 4.0 license** for academic, non-commercial use. Please refer to the license information on the Hugging Face dataset page before using the data.

## Contact

For questions about the dataset, please contact:

```text
Runyu Jiao: rjiao@fbk.eu
Fondazione Bruno Kessler / University of Trento
```