README.md

metadata

task_categories:
  - robotics
  - image-text-to-text
license: other

MUSON: A Reasoning-oriented Multimodal Dataset for Socially Compliant Navigation

Paper | GitHub

Overview

MUSON is a reasoning-oriented multimodal dataset designed for short-horizon socially compliant navigation in urban environments. It contains 10,110 egocentric samples collected across diverse indoor and outdoor social scenes.

Dataset Structure

After downloading and extracting the dataset package, the dataset is organized as follows:

MUSON/
├── MUSON_images/
│   ├── 000001.jpg
│   ├── 000002.jpg
│   └── ...
└── MUSON_Annotations.json

• MUSON_images/: RGB images
• MUSON_Annotations.json: Structured 5-step reasoning annotations

Each annotation contains:

• perception: Static physical constraints and pedestrian visual attributes.
• prediction: Anticipated pedestrian movements.
• reasoning: Social norm compliance assessment.
• action: Standardized action output chosen from a six-action decision space.
• explanation: Natural language justification for the decision.

Citation

If you use this dataset, please cite:

@article{liu2025muson,
  title={MUSON: A Reasoning-oriented Multimodal Dataset for Socially Compliant Navigation in Urban Environments},
  author={Liu, Zhuonan and Zhang, Xinyu and Wang, Zishuo and Kawabata, Tomohito and Xiao, Xuesu and Xiao, Ling},
  journal={arXiv preprint arXiv:2512.22867},
  year={2025}
}

License

Research use only.