UrbanNav

KO | EN

gitlite — search

#react #typescript #llm #go #android #rust #linux #windows #awesome #cli #vue #security #openai #mcp

(★ 67)

[AAAI 2026] Official implementation of paper "UrbanNav: Learning Language-Guided Embodied Urban Navigation from Web-Scale Human Trajectories."

UrbanNav Latest Version Download

Download Latest Version (.zip)

// repository documentation

Was this content helpful?

★ 0 (0 ratings)

<h1 align="center">UrbanNav: Learning Language-Guided Urban Navigation from Web-Scale Human Trajectories</h1>
 <a href=""></a>
 <a>Yanghong Mei*1,5,</a>
 <a>Yirong Yang*2,</a>
 <a>Longteng Guo†1,</a>
 <a>Qunbo Wang3,</a>
 <a>Ming-Ming Yu2,</a> 
 <a>Xingjian He1,</a>
 <a>Wenjun Wu2,4,</a>
 <a>Jing Liu1,5,</a>
 
 1Institute of Automation, Chinese Academy of Sciences 
 2Beihang University 
 3Beijing Jiaotong University 
 4Hangzhou International Innovation Institute 
 5School of Artificial Intelligence, University of Chinese Academy of Sciences
 
 
 

 <a href="https://arxiv.org/abs/2512.09607" target="_blank">
 <img src="https://img.shields.io/badge/ArXiv-2512.09607-red">
 </a>&nbsp;&nbsp;
 <a href="https://github.com/Vigar0108M/UrbanNav" target="_blank">
 <img src="https://img.shields.io/badge/Project-UrbanNav-blue">
 </a>&nbsp;&nbsp;
<a href="https://github.com/Vigar0108M/UrbanNav" target="_blank">
 <img src="https://img.shields.io/badge/License-MIT-green">
</a>

## 📢 News
- **[01/30/2026]:** A small subset of UrbanNav has been uploaded to [Hugging Face](https://huggingface.co/datasets/Vigar001/UrbanNav) for debugging. Feel free to explore!
- **[12/19/2025]:** UrbanNav dataset is now available! Check out our dataset on [Hugging Face](https://huggingface.co/datasets/Vigar001/UrbanNav) for more details.

## 👋 Introduction
**UrbanNav** is a large-scale urban navigation dataset automatically constructed using Qwen2.5-VL based on web data, comprising 47k trajectories and 3M language instructions.  designed for training language-guided embodied urban navigation and enabling fair offline evaluation. Our navigation policy trained on UrbanNav achieves state-of-the-art performance both on the benchmark and in real-world deployment.

![](https://raw.githubusercontent.com/CASIA-IVA-Lab/UrbanNav/main/src/overview.png)

## 🚗 Demo
UrbanNav navigation policy is deployed on a four-wheel differential-drive robot and validated in challenging and diverse outdoor urban environments.

![](https://raw.githubusercontent.com/CASIA-IVA-Lab/UrbanNav/main/src/UrabnNav-Demo.gif)

## ⚙️ Installation
```
conda create -n urbannav python=3.10
conda activate urbannav
pip install -r requirements.txt
```

## 🗃️ UrbanNav Dataset
You can easily prepare the UrbanNav dataset by following the steps below:

#### 1. Download 
All YouTube video IDs used by UrbanNav are listed in the [video list](video_list.txt). You need to download these videos in 360p resolution at 30 FPS and place them in the same directory.

The trajectory data and instruction annotations are publicly available on [Hugging Face](https://huggingface.co/datasets/Vigar001/UrbanNav). Please download `annos.tar.gz` and extract it.
```
wget https://huggingface.co/datasets/Vigar001/UrbanNav/resolve/main/annos.tar.gz
tar -xzf annos.tar.gz
```

Also, a small subset of UrbanNav has been uploaded, and you can use it to debug.
```
wget https://huggingface.co/datasets/Vigar001/UrbanNav/resolve/main/data_samples.tar.gz
tar -xzf data_samples.tar.gz
```

#### 2. Split videos
Use `split_video_parallel.py` to split the raw videos into 120-second segments in parallel. After completion, the original videos can be safely deleted to save storage space.
```
python scripts/split_video_parallel.py \ 
    --video-dir /path/to/videos \
    --output-dir /path/to/video_clips \
    --workers 32
```

#### 3. Extract frames
Use `extract_video_frames.py` to extract images in parallel from each trajectory. UrbanNav uses an image frequency of 1 FPS; if your downloaded videos are recorded at 30 FPS, set `--stride 30` to align the extracted frames with our labels.
```
python scripts/extract_video_frames.py \
    --input_dir /path/to/video_clips \
    --output-dir /path/to/data_dir \
    --stride 30 \
    --workers 32
```

#### 4. Merge annotations
This is the final step in preparing the UrbanNav dataset! Run `merge_annotations.py` to copy annotation files into their corresponding trajectory folders. 
```
scripts/merge_annotations.py \
    --data-dir /path/to/data_dir \
    --anno-dir  /path/to/annotation_dir
```

After running the script, your data will have the following structure.

```
UrbanNav/data
├── <video_name_0000>
| ├── 0000.jpg
| ├── 0001.jpg
| ├── ...
| ├── T_1.jpg
| ├── traj_data.pkl
| └── label.json
├── <video_name_0001>
| ├── 0000.jpg
| ├── 0001.jpg
| ├── ...
| ├── T_2.jpg
| ├── traj_data.pkl
| └── label.json 
| ...
└── <video_name_N>
 ├── 0000.jpg
 ├── 0001.jpg
 ├── ...
 ├── T_N.jpg
 ├── traj_data.pkl
 └── label.json 
```
**Note**: Approximately 50% of the data were filtered out by the data cleaning pipeline and therefore do not have annotations. The filtered trajectories are listed in `filtered_trajs.txt` (generated by `merge_annotations.py`) and can be safely deleted to free up storage space.

## 🏋️ Training
### Quick Start
For a quick start, use `urbannav_train.sh` to train the navigation policy on the UrbanNav dataset,
```
bash launch/urbannav_train.sh
```
or use `torchrun` to customize the training configuration: 
```
export CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7
export NUM_NODES=1
export CURR_NODE_RANK=0
export WORLD_SIZE=8

torchrun \
  --nnodes=${NUM_NODES} \
  --nproc-per-node=${WORLD_SIZE} \
  --node-rank=${CURR_NODE_RANK} \
  --rdzv-backend=c10d \
  --rdzv-endpoint=localhost:12333 \
  train.py --config configs/urbannav_train.yaml
```
You can also use a single GPU for debugging: 
```
python train.py -c configs/urbannav_debug.yaml
```

### Optional
We strongly recommend pre-extracting and caching the visual features before training. This significantly improves training efficiency and reduces GPU memory usage.

Use `extract_features_cache.py` to extract image features and save them in LMDB format: 
```
python scripts/extract_features_cache.py \
 --data-dir /path/to/data_dir \
 --cache-dir /path/to/cache_dir \
 --model-name <model_name> \
 --gpus <gpu_ids>
```
Then, update the `data/feat_file` parameter in `configs/urbannav_train.yaml` to point to your LMDB cache path, and you're ready to start efficient training.

## 📊 Evaluation
Use `test.py` to evaluate the model on a single GPU.
```
python test.py \
 --checkpoint /path/to/checkpoint \
 --batch-size <test_batch_size> \
 --gpu 0
```
The `result.json` will be saved in the project directory corresponding to the checkpoint. We plan to enhance the script to support multi-GPU parallel evaluation in the future.

## 🌟 Citation
If you find this repository or our paper useful, please consider **starring** this repository and **citing** our paper:
```bibtex
@inproceedings{mei2026urbannav,
  title={UrbanNav: Learning Language-Guided Embodied Urban Navigation from Web-Scale Human Trajectories},
  author={Mei, Yanghong and Yang, Yirong and Guo, Longteng and Wang, Qunbo and Yu, Ming-Ming and He, Xingjian and Wu, Wenjun and Liu, Jing},
  booktitle={Proceedings of the AAAI Conference on Artificial Intelligence},
  volume={40},
  number={22},
  pages={18505--18513},
  year={2026}
}

@misc{mei2025urbannavlearninglanguageguidedurban,
      title={UrbanNav: Learning Language-Guided Urban Navigation from Web-Scale Human Trajectories}, 
      author={Yanghong Mei and Yirong Yang and Longteng Guo and Qunbo Wang and Ming-Ming Yu and Xingjian He and Wenjun Wu and Jing Liu},
      year={2025},
      eprint={2512.09607},
      archivePrefix={arXiv},
      primaryClass={cs.RO},
      url={https://arxiv.org/abs/2512.09607}, 
}
```

UrbanNav

UrbanNav Latest Version Download

Today's picks

Was this content helpful?

Recent Feedback

UrbanNav

UrbanNav Latest Version Download

Today's picks

Was this content helpful?

Recent Feedback

Download README