# NavGen

Visual Generative Models as a Scalable Data Engine for Embodied 3D Navigation.

## Description

NavGen contains first-person navigation videos with detailed action descriptions (prompts). Each sample is a short video clip paired with a prompt describing the scene and camera motion.

## Download

### Install HuggingFace Hub

```bash
pip install huggingface_hub
```

### Download via Python

```python
from huggingface_hub import snapshot_download

snapshot_download('xinjiu/NavGen', repo_type='dataset', local_dir='./NavGen')
```

### Download via CLI

```bash
hf download xinjiu/NavGen --repo-type dataset --local-dir ./NavGen
```

### Download via git

```bash
git lfs install
git clone https://huggingface.co/datasets/xinjiu/NavGen
```

## Files

- `NavGen.csv`: video filenames and prompts.
- `navgen_videos_part1.tar.gz` ... `navgen_videos_part4.tar.gz`: the 50,000
  video clips, 12,500 per archive. Each archive is self-contained: extract any
  one of them on its own, no merging required. The file names match the `video`
  column of `NavGen.csv`.

We will release all our datasets once the paper is accepted.