---
title: DeCAL_dataset
canonical_url: "https://www.modelscope.cn/datasets/Aureleo/DeCAL_dataset"
md_url: "https://www.modelscope.cn/datasets/Aureleo/DeCAL_dataset.md"
repository: Aureleo/DeCAL_dataset
last_updated: 2026-09-09
license: "Apache License 2.0"
downloads: 234377
stars: 1
---

# DeCAL_dataset

> DeCAL_dataset - Aureleo 在 ModelScope 开源的数据集。Dataset for DeCAL: Towards Physically-Grounded Dexterous Vision-Language-Action Models via Contact-Aware Latent Co-Imagination.

Aureleo/DeCAL_dataset 是 ModelScope 魔搭社区上的数据集，采用 Apache License 2.0 许可。

- **Repository**: Aureleo/DeCAL_dataset
- **License**: Apache License 2.0
- **Downloads**: 234377
- **Stars**: 1
- **Last updated**: 2026-09-09

Source: https://www.modelscope.cn/datasets/Aureleo/DeCAL_dataset

---

# DeCAL Dataset
<p align="center">
  <a href="https://arxiv.org/abs/2609.09119"><img src="https://img.shields.io/badge/Paper-arXiv-red.svg" alt="Paper"></a>
  <a href="https://www.modelscope.cn/datasets/Aureleo/DeCAL_dataset"><img src="https://img.shields.io/badge/Data-ModelScope-purple" alt="Data"></a>
  <a href="https://aureleopku.github.io/DeCAL/"><img src="https://img.shields.io/badge/Website-Pages-blue.svg" alt="Website"></a>
</p>

Dataset for **DeCAL: Towards Physically-Grounded Dexterous Vision-Language-Action Models via Contact-Aware Latent Co-Imagination**.

This release contains synchronized vision, touch, language, proprioception, and robot trajectories for contact-rich dexterous manipulation. The original DeCAL data has been processed into the unified [Open-X-Tactile (OXT)](https://open-x-tactile.github.io/) format, making it easier to train and evaluate tactile policies alongside other OXT datasets.

> This directory describes the processed OXT release, rather than the original robot logs. The conversion preserves episode boundaries and temporal alignment across modalities.

## Dataset statistics

| Tasks | Trajectories | Frames |
| ---: | ---: | ---: |
| 13 | 1,279 | 1,989,891 |

## Robot setup

The demonstrations were collected on a bimanual platform with:

- two 6-DoF **UR5** robot arms;
- two **Sharpa Wave** five-finger dexterous hands, each with 22 active DoF;
- fingertip tactile sensing on all ten fingers;
- one front RGB camera and two wrist RGB cameras.

The proprioceptive trajectory therefore contains 56 joint-position channels: 6 arm joints and 22 hand joints for each side. See the [Sharpa Wave documentation](https://sharpa-robotics.github.io/sharpa-docs/) for additional hand specifications.

## Modalities

| Modality | Content | Source/native shape |
| --- | --- | --- |
| Language | Time-aligned task instruction | one string per frame |
| RGB vision | Front, left-wrist, and right-wrist views | `480 x 640 x 3` |
| Arm state | Left and right UR5 joint positions | `6 + 6` |
| Hand state | Left and right Sharpa Wave joint positions | `22 + 22` |
| Tactile state | A 6-D contact vector per fingertip | `10 x 6` |
| Tactile deformation | Contact-deformation image for every fingertip | `10 x 240 x 240 x 1` |

The source representation is recorded at **30 Hz**. All image shapes above use `H x W x C`. The converted OXT release provides the synchronized state trajectories required to derive future-state action targets.

## Open-X-Tactile format

The processed data uses OXT-compatible Zarr storage. Each `.zarr` package stores concatenated frame arrays under `data` and episode boundaries under `meta/episode_ends`:

```text
<task-or-batch>.zarr/
├── data/
│   ├── timestamps
│   ├── sub_task_instruction
│   ├── camera_main_rgb
│   ├── left_wrist_camera_rgb
│   ├── right_wrist_camera_rgb
│   ├── left_arm_joints
│   ├── right_arm_joints
│   ├── left_hand_joints
│   ├── right_hand_joints
│   ├── left_hand_joints_idx
│   ├── right_hand_joints_idx
│   └── ...
└── meta/
    └── episode_ends
```

`T` denotes the total number of frames stored in a Zarr package. The main standardized streams follow these shapes:

| OXT field | Typical shape | Description |
| --- | ---: | --- |
| `timestamps` | `(T,)` | Canonical time axis |
| `camera_main_rgb` | `(T, 480, 640, 3)` | Front RGB stream |
| `left_wrist_camera_rgb`, `right_wrist_camera_rgb` | `(T, 480, 640, 3)` | Wrist RGB streams |
| `left_arm_joints`, `right_arm_joints` | `(T, 6)` each | UR5 joint positions |
| `left_hand_joints`, `right_hand_joints` | `(T, 22)` each | Sharpa Wave joint positions |
| `left_hand_joints_idx`, `right_hand_joints_idx` | `(T, 22)` each | Canonical OXT/FAAS joint indices |
| `sub_task_instruction` | `(T,)` | Per-frame language instruction |

### Tactile groups

OXT represents every tactile group with four aligned members:

```text
<side>_tactile_data_<group>
<side>_tactile_area_<group>
<side>_tactile_sensor_<group>
<side>_tactile_type_<group>
```

`data` stores the tactile tensor, `area` identifies the fingertip/contact area, `sensor` is `SharpaWave`, and `type` distinguishes image and state signals. Separate groups preserve deformation images, and per-fingertip contact vectors without conflating their semantics. Exact group suffixes, dtypes, and shapes should be read from the Zarr metadata.

## Loading

```bash
pip install zarr numpy
```

```python
import zarr

root = zarr.open_group('/path/to/task.zarr', mode='r')
data = root['data']
episode_ends = root['meta']['episode_ends'][:]

print('episodes:', len(episode_ends))
for key in sorted(data.keys()):
    array = data[key]
    print(f'{key}: shape={array.shape}, dtype={array.dtype}')
```

To slice one episode:

```python
episode_id = 0
start = 0 if episode_id == 0 else int(episode_ends[episode_id - 1])
end = int(episode_ends[episode_id])

front_rgb = data['camera_main_rgb'][start:end]
instruction = data['sub_task_instruction'][start:end]
```

## Notes

- Modalities belonging to the same frame share the same sequence index after conversion.
- Left/right and finger ordering is preserved across proprioceptive and tactile streams.
- Interpret tactile values together with their OXT `sensor`, `type`, and `area` metadata.
- When deriving future-state actions, never cross an index in `meta/episode_ends`.
- OXT conversion changes the storage schema only. Ownership and licensing remain governed by the original DeCAL dataset release.

## Citation

If you use this dataset, please cite:

```bibtex
@article{fu2026decal,
  title={DeCAL: Towards Physically-Grounded Dexterous Vision-Language-Action Models via Contact-Aware Latent Co-Imagination},
  author={Yankai Fu, Ning Chen, Junkai Zhao, Heng Zhang, Guocai Yao, Pengwei Wang, Zhongyuan Wang, Shanghang Zhang},
  journal={arXiv preprint arXiv:2609.09119},
  year={2026}
}
```

## Acknowledgments

We thank the [Open-X-Tactile](https://open-x-tactile.github.io/) contributors for developing a unified ecosystem for cross-sensor and cross-platform tactile robot learning.
