README.md

---
license: cc-by-nc-sa-4.0
language:
- en
pretty_name: SparseVideoNav Datasets
task_categories:
- robotics
- visual-question-answering
tags:
- embodied-ai
- vision-language-navigation
- robot-navigation
- video
- trajectory
- sparsevideonav
- opendrivelab
configs:
- config_name: bvn
  data_files:
  - split: train
    path: bvn/data.jsonl
- config_name: ifn
  data_files:
  - split: train
    path: ifn/data.jsonl
---

# SparseVideoNav Datasets

![SparseVideoNav dataset overview](assets/dataset_mosaic.png)

This repository contains the real-world navigation datasets released with [OpenDriveLab/SparseVideoNav](https://github.com/OpenDriveLab/SparseVideoNav):

- **BVN**: Beyond-the-View Navigation.
- **IFN**: Instruction-Following Navigation.

Project links:

- Project page: https://opendrivelab.com/SparseVideoNav
- GitHub: https://github.com/OpenDriveLab/SparseVideoNav
- Paper: https://arxiv.org/abs/2602.05827

## Dataset Summary

SparseVideoNav studies real-world vision-language navigation with sparse future video generation. The datasets contain language instructions, RGB frame sequences, and low-level navigation actions. The number of actions matches the number of RGB frames for every released episode.

Due to policy restrictions, this public release contains the policy-compliant subset of the collected real-world data. At the released sampling rate of **4 fps**, the open-sourced subset contains **about 100 hours** of navigation data.

| Subset | Episodes | RGB frames | Duration @ 4 fps | Task |
| --- | ---: | ---: | ---: | --- |
| `bvn` | 5,417 | 695,373 | 48.29 h | Beyond-the-View Navigation |
| `ifn` | 5,625 | 712,290 | 49.46 h | Instruction-Following Navigation |
| **Total** | **11,042** | **1,407,663** | **97.75 h** | - |

## Repository Structure

Images are stored in compressed tar shards to avoid hundreds of thousands of small files in the Hugging Face repository. Each shard preserves the original relative paths.

```text
.
├── README.md
├── assets/
│   └── dataset_mosaic.png
├── bvn/
│   ├── annotations.json
│   ├── data.jsonl
│   ├── merge_info.json
│   ├── shard_manifest.jsonl
│   └── shards/
│       ├── bvn-00000.tar.zst
│       └── ...
└── ifn/
    ├── annotations.json
    ├── data.jsonl
    ├── merge_info.json
    ├── shard_manifest.jsonl
    └── shards/
        ├── ifn-00000.tar.zst
        └── ...
```

## Data Format

Each line in `bvn/data.jsonl` or `ifn/data.jsonl` is an episode-level JSON object.

| Field | Type | Description |
| --- | --- | --- |
| `dataset` | string | Dataset subset name, either `bvn` or `ifn`. |
| `subset` | string | Release subset marker. The current public release uses `main`. |
| `episode_id` | string | Unique episode identifier. |
| `instruction` | string | Primary natural-language navigation instruction. |
| `instructions` | list[string] | Instruction list. Current records contain one instruction. |
| `task_type` | string | Task label, e.g. `beyond_the_view_navigation` or `instruction_following_navigation`. |
| `split` | string | Dataset split. Current release uses `train`. |
| `image_dir` | string | Original relative episode image directory before sharding. |
| `rgb_dir` | string | Original relative RGB frame directory. |
| `num_frames` | integer | Number of RGB frames in the episode. |
| `num_actions` | integer | Number of low-level actions. This matches `num_frames`. |
| `actions` | list[object] | Per-frame low-level navigation actions. Each action has `dx`, `dy`, and `dyaw`. |
| `annotation_file` | string | Annotation file path in this repository. |
| `frame_files` | list[string] | RGB frame paths inside the shard after extraction. |
| `asset_dirs` | list[string] | Original episode asset directories included in the shard. |
| `shard` | string | Tar shard containing the episode frames. |

Each action object contains:

| Field | Type | Description |
| --- | --- | --- |
| `dx` | float | Relative forward/backward displacement for the corresponding step. |
| `dy` | float | Relative lateral displacement for the corresponding step. |
| `dyaw` | float | Relative yaw change for the corresponding step. |

Example:

```json
{
  "dataset": "ifn",
  "episode_id": "36e980e5a1ff4817",
  "instruction": "please go along with the rail until you are near by a red cone.",
  "num_frames": 177,
  "num_actions": 177,
  "rgb_dir": "images/0e0ec1bdcc61a4ef/rgb",
  "frame_files": ["images/0e0ec1bdcc61a4ef/rgb/001.jpg"],
  "shard": "ifn/shards/ifn-00000.tar.zst"
}
```

`annotations.json` stores the original annotation records with the core fields `id`, `video`, `actions`, and `instructions`. `shard_manifest.jsonl` stores shard-level metadata, including the shard path, episode ids, raw byte size, compressed byte size, frame count, and file count.

## Usage

Load episode metadata with Hugging Face Datasets:

```python
from datasets import load_dataset

bvn = load_dataset("OpenDriveLab/SparseVideoNav", "bvn")
ifn = load_dataset("OpenDriveLab/SparseVideoNav", "ifn")
```

Download and inspect shards:

```bash
tar -I zstd -tf ifn/shards/ifn-00000.tar.zst | head
tar -I zstd -xf ifn/shards/ifn-00000.tar.zst
```

After extraction, `frame_files` resolve to paths such as:

```text
images/<episode_dir>/rgb/001.jpg
```

## License

The dataset is released under CC BY-NC-SA 4.0.

## Citation

```bibtex
@article{zhang2026sparse,
  title={Sparse Video Generation Propels Real-World Beyond-the-View Vision-Language Navigation},
  author={Zhang, Hai and Liang, Siqi and Chen, Li and Li, Yuxian and Xu, Yukuan and Zhong, Yichao and Zhang, Fu and Li, Hongyang},
  journal={arXiv preprint arXiv:2602.05827},
  year={2026}
}
```