README.md

---
license: mit
task_categories:
- reinforcement-learning
tags:
- robotics
- dexterous-manipulation
- robot-learning
- imitation-learning
- offline-reinforcement-learning
- shadow-hand
- mujoco
- hdf5
- cvpr-2025
pretty_name: DexHandDiff Manipulate Block
size_categories:
- 1K<n<10K
---

# DexHandDiff Manipulate Block Dataset

This dataset contains expert demonstration trajectories for the **Manipulate Block / Block Rotate-Z** task used in [DexHandDiff: Interaction-aware Diffusion Planning for Adaptive Dexterous Manipulation](https://arxiv.org/abs/2411.18562), accepted at CVPR 2025.

The task evaluates in-hand dexterous manipulation with a simulated Shadow Hand. The hand must reorient a block to a target pose, with particular emphasis on rotation about the z-axis while maintaining alignment in the remaining orientation dimensions. The demonstrations were used to train and evaluate diffusion-based planners for contact-rich, goal-adaptive dexterous manipulation.

## Dataset Description

- **Task:** Manipulate Block / Block Rotate-Z
- **Environment:** `HandManipulateBlockRotateXYZ-v1`
- **Embodiment:** 24-joint Shadow Hand, with the full environment containing up to 30 degrees of freedom
- **Number of trajectories:** 5,000 expert trajectories
- **Number of transitions:** 123,711
- **Collection method:** Truncated Quantile Critics with Hindsight Experience Replay (TQC+HER)
- **Modality:** state-action robot trajectories
- **Primary file:** `manipulate_block_5000.h5`
- **Associated paper:** [arXiv:2411.18562](https://arxiv.org/abs/2411.18562)
- **Project page:** [dexdiffuser.github.io](https://dexdiffuser.github.io/)
- **Code:** [github.com/Liang-ZX/DexHandDiff](https://github.com/Liang-ZX/DexHandDiff)

The Adroit tasks in the paper use teleoperated demonstrations from D4RL. The Shadow Hand environment does not provide demonstration data for this task, so we collected these 5,000 expert trajectories using TQC+HER.

## Task and Evaluation Protocol

In the Block Rotate-Z task, the dexterous hand continuously reorients a block toward a target orientation represented in quaternion space. The task requires coordinated multi-finger control and contact-rich in-hand adjustment.

The goal-adaptability experiment in the paper uses a half-side split: the training goals have positive target yaw angles, while evaluation includes negative target yaw angles. This tests whether a planner can adapt beyond the goal distribution represented during training.

The paper reports the following success rates for DexHandDiff:

| Evaluation task | Success rate |
| --- | ---: |
| Rotate-Z | 50.0 ± 8.2% |
| Half-side Rotate-Z | 36.7 ± 4.7% |

These values are method results reported in the paper and should not be interpreted as intrinsic properties of the dataset.

## Files and Data Format

The dataset is distributed as a single HDF5 file:

```text
manipulate_block_5000.h5
```

The archive contains **5,000 expert trajectories** and **123,711 transition steps** collected in `HandManipulateBlockRotateXYZ-v1`.

| Field | Shape | Dtype | Description |
| --- | --- | --- | --- |
| `observation` | `(123711, 1, 61)` | `float32` | Environment observations |
| `action` | `(123711, 1, 20)` | `float32` | Shadow Hand actions |
| `reward` | `(123711, 1)` | `float32` | Sparse task rewards |
| `terminal` | `(123711, 1)` | `bool` | Stored terminal indicators |
| `info/achieved_goal` | `(123711, 1, 7)` | `float32` | Achieved goal states |
| `info/desired_goal` | `(123711, 1, 7)` | `float32` | Desired goal states |
| `info/is_success` | `(123711,)` | `float32` | Task-success indicators |
| `info/object_qvel` | `(123711, 6)` | `float64` | Object velocities |
| `info/robot_qpos` | `(123711, 24)` | `float64` | Shadow Hand joint positions |
| `info/robot_qvel` | `(123711, 24)` | `float64` | Shadow Hand joint velocities |
| `info/timeout` | `(123711,)` | `bool` | Stored timeout indicators |

The sparse reward is `-1` before task completion and `0` upon success. The file contains 5,000 positive `info/is_success` markers, corresponding to the 5,000 collected expert trajectories. The stored `terminal` and `info/timeout` arrays do not mark every trajectory boundary; use `info/is_success` to recover the successful trajectory boundaries in collection order.

The HDF5 root attributes are:

| Attribute | Value |
| --- | --- |
| `env_name` | `HandManipulateBlockRotateXYZ-v1` |
| `total_episodes` | `5000` |

To inspect the groups, fields, shapes, and dtypes programmatically:

```python
import h5py

path = "manipulate_block_5000.h5"

with h5py.File(path, "r") as dataset:
    def describe(name, obj):
        if isinstance(obj, h5py.Dataset):
            print(f"{name}: shape={obj.shape}, dtype={obj.dtype}")
        else:
            print(f"{name}/")

    dataset.visititems(describe)
```

After the file is available on the Hub, it can be downloaded programmatically with:

```python
from huggingface_hub import hf_hub_download

path = hf_hub_download(
    repo_id="Liang-ZX/DexHandDiff-ManipulateBlock",
    filename="manipulate_block_5000.h5",
    repo_type="dataset",
)
```

Because HDF5 is a hierarchical container rather than a table-oriented format, the Hugging Face Dataset Viewer may not render the trajectories directly. Use `h5py` or the data-loading utilities in the DexHandDiff codebase to access the data.

## Intended Uses

This dataset is intended for research on:

- dexterous and contact-rich robot manipulation;
- imitation learning and offline reinforcement learning;
- diffusion-based planning and policy learning;
- goal-conditioned and goal-adaptive control;
- state-action dynamics modeling; and
- in-hand object reorientation.

## Limitations

- The trajectories were collected in simulation and may not transfer directly to a physical Shadow Hand without additional adaptation.
- The dataset covers a single block-reorientation environment and should not be treated as a broad benchmark of general dexterous manipulation.
- Performance may depend on the exact simulator version, observation/action definitions, preprocessing, and evaluation protocol.
- The goal distribution used for training does not represent every possible block orientation.

## License

This dataset is released under the MIT License. Users are also responsible for complying with the licenses of the underlying simulator and software dependencies.

## Citation

If you use this dataset, please cite:

```bibtex
@InProceedings{Liang_2025_CVPR,
    author    = {Liang, Zhixuan and Mu, Yao and Wang, Yixiao and Chen, Tianxing and Shao, Wenqi and Zhan, Wei and Tomizuka, Masayoshi and Luo, Ping and Ding, Mingyu},
    title     = {DexHandDiff: Interaction-aware Diffusion Planning for Adaptive Dexterous Manipulation},
    booktitle = {Proceedings of the Computer Vision and Pattern Recognition Conference (CVPR)},
    month     = {June},
    year      = {2025},
    pages     = {1745--1755}
}
```

## Questions

For questions about the dataset, please open an issue in the [DexHandDiff repository](https://github.com/Liang-ZX/DexHandDiff/issues) or start a discussion in this Hugging Face dataset repository.