---
task_categories:
- robotics
tags:
- human-robot-interaction
- multi-agent
- embodied-ai
- reinforcement-learning
dataset_info:
  features:
  - name: map_name
    dtype: string
  - name: trajectory_id
    dtype: int64
  - name: steps_data
    dtype: string
  - name: num_steps
    dtype: int64
  splits:
  - name: train
    num_bytes: 562424179
    num_examples: 912
  - name: evaluation
    num_bytes: 54684865
    num_examples: 96
  - name: all_data
    num_bytes: 617109044
    num_examples: 1008
  download_size: 261118576
  dataset_size: 1234218088
configs:
- config_name: default
  data_files:
  - split: train
    path: data/train-*
  - split: evaluation
    path: data/evaluation-*
  - split: all_data
    path: data/all_data-*
---

# Moving Out: Physically-grounded Human-AI Collaboration Dataset

This repository contains the human-human interaction dataset for the paper [Moving Out: Physically-grounded Human-AI Collaboration](https://huggingface.co/papers/2507.18623).

The **Moving Out** benchmark focuses on human-AI collaboration in physical environments, specifically for embodied agents (e.g., robots). It introduces a new benchmark that resembles a wide range of collaboration modes affected by physical attributes and constraints, such as moving heavy items together and maintaining consistent actions to move a big item around a corner. The dataset includes human-human interaction data collected for two tasks, designed to evaluate models' abilities to adapt to diverse human behaviors and unseen physical attributes in continuous state-action spaces with constrained dynamics.

**Project Page:** [https://live-robotics-uva.github.io/movingout_ai/](https://live-robotics-uva.github.io/movingout_ai/)

**Code:** [https://github.com/live-robotics-uva/moving_out_ai](https://github.com/live-robotics-uva/moving_out_ai)

## Sample Usage

You can visualize trajectories from this dataset using the `dataset_to_video.py` script provided in the associated GitHub repository.

**Save trajectory as MP4:**

```bash
python dataset_to_video.py -f ShuaKang/movingout_task2 -m HandOff -t 4 -v video
```

**Show trajectory in popup window:**

```bash
python dataset_to_video.py -f ShuaKang/movingout_task2 -m HandOff -t 4 -v human
```

Use `-m` for map name and `-t` for trajectory ID.

## Citation

If you use this dataset in your research, please cite the original paper:

```bibtex
@article{,
      title={Moving Out: Physically-grounded Human-AI Collaboration},
      author={Shua Kang, Wenxin Xia, Ziqi Li, Mingxing Yuan, Xufan Wu, Min Bai, Boyi Liu, David Abel, Wenhao Zhang},
      journal={arXiv preprint arXiv:2507.18623},
      year={2025},
      url={https://huggingface.co/papers/2507.18623}
}
```