---
license: mit
task_categories:
- robotics
---

# MetaFine Dataset

This dataset is a companion release for the paper [Beyond Binary Success: A Diagnostic Meta-Evaluation Framework for Fine-Grained Manipulation](https://huggingface.co/papers/2605.19986).

- **Project Page:** [https://metafine.github.io/](https://metafine.github.io/)
- **GitHub Repository:** [https://github.com/Hiangx-robotics/MetaFine](https://github.com/Hiangx-robotics/MetaFine)

MetaFine is a diagnostic meta-evaluation framework for fine-grained robotic manipulation. Instead of a single binary success rate, it disentangles manipulation competency along three fundamental axes: **understanding**, **perception**, and **behavior**.

## Dataset Structure

The repository contains two main parts:

- **Fine-grained annotated simulation assets**: 40+ URDF-based simulation objects (including a PartNet-Mobility subset and custom URDFs) with fine-grained part annotations and related interaction metadata (`capabilities.json`).
- **Fine-grained task data**: Task specifications (YAML) used for training and testing fine-grained manipulation policies across various environments.

## Usage

You can download the assets and configurations using the Hugging Face CLI:

```bash
huggingface-cli download hiangx/MetaFine --repo-type dataset
```

Place the unpacked `assets/` and `configs/` directories next to your local MetaFine repository root to use them with the platform.

## Citation

```bibtex
@article{xu2026metafine,
  title   = {Beyond Binary Success: A Diagnostic Meta-Evaluation Framework for Fine-Grained Manipulation},
  author  = {Xu, He-Yang and Zhang, Pengyuan and Ge, Zongyuan and Hao, Xiaoshuai and Belongie, Serge and Geng, Xin and Peng, Yuxin and Wei, Xiu-Shen},
  journal = {arXiv preprint arXiv:2605.19986},
  year    = {2026}
}
```