Dataset Card for Dataset Name

MetaFold Dataset is a point-cloud trajectory dataset designed for multi-category garment folding tasks in robotic manipulation.

Dataset Details

Dataset Description

• Curated by: Haonan Chen, Junxiao Li, Ruihai Wu, Yiwei Liu, Yiwen Hou, Zhixuan Xu, Jingxiang Guo, Chongkai Gao, Zhenyu Wei, Shensi Xu, Jiaqi Huang, Lin Shao
• License: CC BY-NC-ND 4.0

Dataset Sources [optional]

• Repository: https://huggingface.co/datasets/chenhn02/MetaFold
• Paper: MetaFold: Language-Guided Multi-Category Garment Folding Framework via Trajectory Generation and Foundation Model
• Website: https://meta-fold.github.io/

Dataset Structure

The dataset categories is inherited from ClothesNet. Each garment category has a corresponding name (e.g., DLG stands for Dress-Long-Gallus).

Four Folding Types in MetaFold

1. No-Sleeve Folding

   • Categories:
     • DLG (Dress-Long-Gallus)
     • DLNS (Dress-Long-NoSleeve)
     • DLT (Dress-Long-Tube)
     • DSNS (Dress-Short-NoSleeve)
     • SL (Skirt-Long)
     • SS (Skirt-Short)
     • TCNC (Top-Collar-NoSleeve-FrontClose)
     • TCNO (Top-Collar-NoSleeve-FrontOpen)
     • THNC (Top-Hooded-NoSleeve-FrontClose)
     • TNNC (Top-NoCollar-NoSleeve-FrontClose)
   • Total garments: 666
   • Folding procedure (1 stage):
   • 1. Fold from bottom to the top.

2. Short-Sleeve Folding

   • Categories:
     • DLSS (Dress-Long-ShortSleeve)
     • DSSS (Dress-Short-ShortSleeve)
     • TNSC (Top-NoCollar-ShortSleeve-FrontClose)
     • TCSC (Top-Collar-ShortSleeve-FrontClose)
   • Total garments: 121
   • Folding procedure (3 stages):
     1. Fold the left sleeve.
     2. Fold the right sleeve.
     3. Fold from the bottom hem upward.

3. Long-Sleeve Folding

   • Categories:
     • DLLS (Dress-Long-LongSleeve)
     • DSLS (Dress-Short-LongSleeve)
     • TCLC (Top-Collar-LongSleeve-FrontClose)
     • TCLO (Top-Collar-LongSleeve-FrontOpen)
     • THLC (Top-Hooded-LongSleeve-FrontClose)
     • THLO (Top-Hooded-LongSleeve-FrontOpen)
     • TNLC (Top-NoCollar-LongSleeve-FrontClose)
   • Total garments: 146
   • Folding procedure (3 stages):
     1. Fold the left sleeve.
     2. Fold the right sleeve.
     3. Fold from the bottom hem upward.

4. Pants Folding

   • Categories:
     • PS (Pants-Short)
     • PL (Pants-Long)
   • Total garments: 277
   • Folding procedure (2 stages):
     1. Fold the left pant leg over the right pant leg.
     2. Fold from the waistband down toward the pant legs.

Total Counts

• No-Sleeve: 666 garments
• Short-Sleeve: 121 garments
• Long-Sleeve: 146 garments
• Pants: 277 garments
• Overall Total: 1,210 garments

Filename Convention

Each data file follows the pattern: <Category>_<GarmentName>_<FoldStage>

• Category: Garment category code (e.g., DLG = Dress-Long-Gallus).
• GarmentName: Unique identifier for the specific garment model (e.g., Dress032_1, 010).
• FoldStage: Zero-based index indicating which folding stage (e.g., action0, action1, …).

Examples

1. DLG_Dress032_1_action0

   • Category: DLG (Dress-Long-Gallus → No-Sleeve)
   • GarmentName: Dress032_1
   • FoldStage: action0 (since No-Sleeve has only one stage: fold from bottom to top)
   • Link: DLG_Dress032_1_action0

2. TNLC_010_action1

   • Category: TNLC (Tops-NoCollar-LongSleeve-Close → Long-Sleeve)
   • GarmentName: 010
   • FoldStage: action1 (second stage: fold the right sleeve)
   • Link: TNLC_010_action1

Directory Content

Each Folding Trajectory directory contains:

• Mesh files from frame_00.obj to frame_20.obj.
• initial.obj is identical to frame_00.obj.
• keypoints_idx.npy (specifies the folding start and end points for each stage: [start_0, end_0, start_1, end_1]).

If you need the point cloud, simply extract the vertex coordinates from the mesh files.

Dataset Creation

Curation Rationale

Current garment‐folding datasets are very limited, and there is a lack of data in the deformable‐object manipulation domain. To bridge this gap and benefit the entire embodied‐intelligence community, we hereby release the dataset constructed in this work.

Data Collection and Processing

We use data from ClothesNet and employ DiffClothAI as the simulator. We heuristically specify the folding start and end points, then simulate the folding process within the simulator. For the detailed procedure, please refer to our paper MetaFold. The specific keypoints we designated are stored in each trajectory directory under keypoints_idx.npy.

Citation

BibTeX:

@misc{chen2025metafoldlanguageguidedmulticategorygarment,
  title={MetaFold: Language-Guided Multi-Category Garment Folding Framework via Trajectory Generation and Foundation Model}, 
  author={Haonan Chen and Junxiao Li and Ruihai Wu and Yiwei Liu and Yiwen Hou and Zhixuan Xu and Jingxiang Guo and Chongkai Gao and Zhenyu Wei and Shensi Xu and Jiaqi Huang and Lin Shao},
  year={2025},
  eprint={2503.08372},
  archivePrefix={arXiv},
  primaryClass={cs.RO},
  url={https://arxiv.org/abs/2503.08372}, 
}

Dataset Card Contact

First Author: Haonan Chen Corresponding Author: Lin Shao