Leiyao-Cui/XieNet

Dataset Card for XieNet

This is the repaired version of GAPartNet dataset, which we use as the simulation dataset for Vi-TacMan.

Description

We identified numerous object meshes in the original dataset that lack proper cap geometry, so we manually repaired these meshes to ensure completeness. The following images (object id: 47296) exemplify the type of geometric defects found and our corrections:

<div align="center"> <table> <tr> <td align="center"> <img src="figure/gapartnet47296.png" width="300" alt="GAPartNet Original"/> <br/> GAPartNet (Original) </td> <td align="center"> <img src="figure/xienet47296.png" width="300" alt="XieNet Repaired"/> <br/> XieNet (Repaired) </td> </tr> </table> </div>

We also provide the data generation code, which can be used to reproduce the simulated data presented in our paper Vi-TacMan.

We sincerely thank the previous works (SAPIEN, PartNet, GAPartNet) and hope our repaired dataset can help advance this community.

Usage

Installation

First, install the required dependencies:

pip install -r requirements.txt

Requirements:

• —Python 3.10
• —SAPIEN 3.0.1

Data Generation

The main script main.py generates simulated data by rendering articulated objects from multiple camera viewpoints with different articulation states.

Basic Usage

python main.py \
    --data_root_dir /path/to/XieNet \
    --save_dir /path/to/output/directory

Full Command Line Options

python main.py \
    --data_root_dir /path/to/XieNet/dataset \  # Path to the XieNet dataset root
    --save_dir /path/to/output/directory \     # Output directory for rendered data
    --seed 42 \                                # Random seed (default: 42)
    --render_width 640 \                       # Render width (default: 640)
    --render_height 576 \                      # Render height (default: 576)
    --fovy 65.0 \                              # Field of view in degrees (default: 65.0)
    --near 0.01 \                              # Near clipping plane (default: 0.01)
    --far 4.0 \                                # Far clipping plane (default: 4.0)
    --enable_rt \                              # Enable ray tracing (optional)
    --min_movable_area 4096 \                  # Minimum area for movable parts (default: 4096)
    --max_flow_dist 0.1 \                      # Maximum flow distance (default: 0.1)
    --save_vis                                 # Save visualization images (default: True)

Supported Object Categories

The data generation focuses on the following articulated object categories, for which we provide repaired meshes:

• —Dishwasher
• —Door
• —Microwave
• —Oven
• —Refrigerator
• —Safe
• —StorageFurniture
• —Table
• —Toilet
• —TrashCan
• —WashingMachine

Output Data Format

For each object and camera viewpoint, the script generates:

• —pcd_camera.npy: Structured numpy array containing:
• —point: 3D point coordinates in camera frame
• —rgb: RGB color values
• —articulation_flow: 3D flow vectors for articulation motion
• —mask_holdable: Binary mask for holdable parts
• —mask_movable: Binary mask for movable parts
• —mask_ground: Binary mask for ground plane
• —camera_pose.txt: 4x4 camera pose matrix
• —camera_intrinsics.txt: 3x3 camera intrinsic matrix
• —vis/ folder (if --save_vis is enabled): Visualization images including color, depth, masks, and flow visualizations

Citation

If you find this dataset beneficial, please cite our research paper as follows:

@inproceedings{cui2026vitacman,
  title = {Vi-{T}ac{M}an: Articulated Object Manipulation via Vision and Touch},
  author = {Cui, Leiyao and Zhao, Zihang and Xie, Sirui and Zhang, Wenhuan and Han, Zhi and Zhu, Yixin},
  booktitle = {IEEE International Conference on Robotics and Automation (ICRA)},
  year = {2026},
  organization = {IEEE}    
}