---
language:
- en
license: mit
size_categories:
- 10K<n<100K
task_categories:
- video-text-to-text
tags:
- GUI
- CUA
- Agents
- action prediction
- multimodal
- computer-use
- video-demonstrations
- desktop-automation
---

<p align="center">
  <img src="assets/cua-suite-logo.png" alt="CUA-Suite Logo" width="120"/>
</p>

<h1 align="center"><font size="7">VideoCUA</font></h1>

<p align="center">
  <strong>The largest open, human annotated video corpus for desktop computer use</strong><br>
  Part of <a href="https://cua-suite.github.io/">CUA-Suite</a>: Massive Human-annotated Video Demonstrations for Computer-Use Agents
</p>

<p align="center">
  <a href="https://arxiv.org/abs/2603.24440">Paper</a> •
  <a href="https://cua-suite.github.io/">Project Page</a> •
  <a href="https://github.com/ServiceNow/GroundCUA/tree/main/VideoCUA">GitHub</a> •
  <a href="https://uivision.github.io/">UI-Vision</a> •
  <a href="https://groundcua.github.io/">GroundCUA</a>
</p>

<p align="center">
  <img src="assets/cua-suite-teaser.png" alt="CUA-Suite Teaser" width="100%"/>
</p>

## Overview

**VideoCUA** is the largest open expert video corpus for desktop computer use, comprising **~10K tasks**, **55 hours** of continuous 30 fps screen recordings, and **6 million frames** across **87 professional desktop applications** spanning 12 categories.

Unlike sparse screenshot datasets, VideoCUA preserves the full temporal dynamics of human interaction — every mouse movement, click, drag, scroll, and keystroke is logged with millisecond precision alongside continuous video. This enables research in action prediction, imitation learning, visual world models, and video-based reward modeling.

VideoCUA is part of [CUA-Suite](https://cua-suite.github.io/), a unified ecosystem that also includes:

- [**UI-Vision**](https://uivision.github.io/) — A desktop-centric benchmark evaluating element grounding, layout understanding, and action prediction.
- [**GroundCUA**](https://groundcua.github.io/) — A large-scale pixel-precise UI grounding dataset with 5M+ human-verified element annotations.

## Usage

To process the raw video data and action logs into trajectories for training or evaluation, you can use the synthesis pipeline provided in the [GitHub repository](https://github.com/ServiceNow/GroundCUA/tree/main/VideoCUA).

### 1. Download & Extract
```bash
bash download_data.sh --repo ServiceNow/VideoCUA --output_dir ./VideoCUA
```

### 2. Convert to Trace Format
To extract video frames at each action timestamp and produce standardized trajectories:
```bash
python convert_videocua.py \
    --data_dir ./VideoCUA/data \
    --output_dir ./videocua_processed \
    --num_workers 4
```

### 3. Generate CoT Annotations
```bash
python gen_cot.py \
    --task_list_path ./videocua_processed/task_list.json \
    --model claude-sonnet-4.5 \
    --num_threads 4 \
    --suffix cot_v1
```

## Repository Structure

```
.
├── assets/
│   ├── cua-suite-logo.png
│   └── cua-suite-teaser.png
├── raw_data/                         # One zip per application (87 total)
│   ├── 7-Zip.zip
│   ├── Affine.zip
│   ├── Anki.zip
│   ├── ...
│   └── draw.io.zip
└── README.md
```

## Data Format

Each application zip in `raw_data/` contains multiple task folders identified by numeric task IDs. Each task folder has the following structure:

```
<task_id>/
├── action_log.json          # Task metadata and timestamped actions
└── video/
    ├── video.mp4            # Continuous 30 fps screen recording (1920×1080)
    └── video_metadata.json  # Video properties (fps, duration, resolution, etc.)
```

### `action_log.json`

```json
{
  "task_id": 45525,
  "task_instruction": "Open test.7z present in archive it and see the contents",
  "platform": "7-Zip",
  "action_log": [
    {
      "action_type": "CLICK",
      "timestamp": 2.581,
      "action_params": {
        "x": 47,
        "y": 242,
        "numClicks": 2
      },
      "groundcua_id": "9a7daeed..."
    }
  ]
}
```

Each action entry includes a `groundcua_id` field — this is the unique identifier for the corresponding screenshot in the [GroundCUA](https://huggingface.co/datasets/ServiceNow/GroundCUA) repository. Using this ID, you can look up the fully annotated screenshot in GroundCUA, linking the video action trajectory to dense UI grounding annotations.

## Citation

If you find VideoCUA or any of the other works in CUA-Suite useful for your research, please cite our works:

```bibtex
@inproceedings{
    jian2026cuasuite,
    title={{CUA}-Suite: Massive Human-annotated Video Demonstrations for Computer-Use Agents},
    author={Xiangru Jian and Shravan Nayak and Kevin Qinghong Lin and Aarash Feizi and Kaixin Li and Patrice Bechard and Spandana Gella and Sai Rajeswar},
    booktitle={ICLR 2026 Workshop on Lifelong Agents: Learning, Aligning, Evolving},
    year={2026},
    url={https://openreview.net/forum?id=IgTUGrZfMr}
}

@inproceedings{
    feizi2026grounding,
    title={Grounding Computer Use Agents on Human Demonstrations},
    author={Aarash Feizi and Shravan Nayak and Xiangru Jian and Kevin Qinghong Lin and Kaixin Li and Rabiul Awal and Xing Han L{\`u} and Johan Obando-Ceron and Juan A. Rodriguez and Nicolas Chapados and David Vazquez and Adriana Romero-Soriano and Reihaneh Rabbany and Perouz Taslakian and Christopher Pal and Spandana Gella and Sai Rajeswar},
    booktitle={The Fourteenth International Conference on Learning Representations},
    year={2026},
    url={https://openreview.net/forum?id=9WiPZy3Kro}
}

@inproceedings{
    nayak2025uivision,
    title={{UI}-Vision: A Desktop-centric {GUI} Benchmark for Visual Perception and Interaction},
    author={Shravan Nayak and Xiangru Jian and Kevin Qinghong Lin and Juan A. Rodriguez and Montek Kalsi and Nicolas Chapados and M. Tamer {\"O}zsu and Aishwarya Agrawal and David Vazquez and Christopher Pal and Perouz Taslakian and Spandana Gella and Sai Rajeswar},
    booktitle={Forty-second International Conference on Machine Learning},
    year={2025},
    url={https://openreview.net/forum?id=5Rtj4mYH1C}
}
```

## License

This dataset is released under the [MIT License](LICENSE).