HR-Motion

HR-Motion is a dataset for hierarchical motion relation prediction in 3D CAD assemblies.

This dataset is introduced in our paper:

HMRNet: Hierarchical Motion Relation Learning for 3D CAD Assemblies

Code repository:

https://github.com/shisyy/HMRNet

Dataset Description

HR-Motion is constructed from publicly available and free 3D CAD assembly models collected from online repositories. We manually annotated hierarchical motion relations for these CAD assemblies.

The dataset contains STEP/BRep CAD assembly models and structured annotations for solid-pair hierarchical motion relations.

HR-Motion defines five base motion relation types between solids:

• set
• rotation
• translation
• screw
• others

For rotation, translation, and screw relations, parent-child directional annotations are provided to describe hierarchical motion dependencies.

Dataset Statistics

Split	Relations	Set	Rotation	Translation	Screw	Others
Train	45,263	24,438	1,311	1,822	378	17,314
Test	15,724	9,698	494	486	102	4,944
Total	60,987	34,136	1,805	2,308	480	22,258

Annotation Format

Each binary relation between two solids is represented by a 14-dimensional annotation vector, including:

• solid indices
• relation type
• parent-child direction
• joint position
• joint direction
• joint state
• motion range bounds
• screw ratio

In the current benchmark setting, the main task focuses on hierarchical motion relation prediction using the solid indices, relation type, and parent-child direction.

Files

• HR-Motion.7z: compressed archive containing the HR-Motion dataset, including STEP/BRep CAD assembly models and hierarchical motion relation annotations.

Because the dataset is distributed as a compressed .7z archive, the Hugging Face Dataset Viewer may not be available. Please download and extract the archive manually.

Usage

Download and extract the dataset:

7z x HR-Motion.7z

Then place the extracted data under the data directory of the HMRNet codebase. Please refer to the official repository for environment setup, data preparation, and training instructions:

https://github.com/shisyy/HMRNet

License

The HR-Motion annotations are released under CC BY-NC 4.0 for non-commercial research use.

The CAD models were collected from publicly available and free online sources. Users should comply with the original licenses or terms of the corresponding CAD model sources.

Citation

If you use HR-Motion or HMRNet in your research, please cite our paper:

@misc{shi2026hmrnet,
  title={HMRNet: Hierarchical Motion Relation Learning for 3D CAD Assemblies},
  author={Shi, Yingying and Zhao, Jiaxuan and Zhou, Shengdi and Tao, Ye and Zhou, Zhong and Zhou, Bin},
  year={2026}
}

Contact

For questions about the dataset or code, please contact the authors or open an issue in the GitHub repository.