📢 News

• [Latest] UAV-Flow real-world trajectories and UAV-Flow-Sim simulation trajectories are available on Hugging Face.
• [Latest] UAV-Flow-Eval and the OpenVLA-UAV training/evaluation code are included in this repository.
• [Coming Soon] Pi-0-UAV support will be released later.

💡 Highlights

• Real and simulated UAV data. UAV-Flow contains real-world UAV trajectories, while UAV-Flow-Sim provides simulation trajectories for scalable training and evaluation.
• Instruction-conditioned evaluation. UAV-Flow-Eval runs UAV control tasks in UnrealZoo Gym and evaluates generated trajectories with nDTW.
• OpenVLA-UAV baseline. The repository includes OpenVLA-based finetuning code, a Flask inference server, and a released OpenVLA-UAV checkpoint.

🛠️ Usage

Installation

This repository contains two main Python environments: one for OpenVLA-UAV training/inference and one for UAV-Flow-Eval.

Set up OpenVLA-UAV:

cd OpenVLA-UAV
conda create -n openvla python=3.10 -y
conda activate openvla

conda install pytorch torchvision torchaudio pytorch-cuda=12.4 -c pytorch -c nvidia -y
pip install -e .

pip install packaging ninja
ninja --version
pip install "flash-attn==2.5.5" --no-build-isolation

Set up UAV-Flow-Eval:

cd UAV-Flow-Eval
conda create -n unrealcv python=3.11 -y
conda activate unrealcv
pip install -e .

Data Preparation

Download the datasets:

• UAV-Flow: real-world trajectories.
• UAV-Flow-Sim: simulation trajectories.

If you need to convert parquet files into the folder format used by the training code, update the local paths in dataset_tools/prepare_data.py, then run:

cd dataset_tools
python prepare_data.py

Model Preparation

For finetuning, download the pretrained OpenVLA checkpoint.

For evaluation, download the released OpenVLA-UAV checkpoint, then set model_path in OpenVLA-UAV/vla-scripts/openvla_act.py:

cfg = {
    "gpu_id": 0,
    "model_path": "/path/to/your/finetuned_model",
    "http_port": 5007,
    "unnorm_key": "sim",
    "do_sample": False,
}

Training

Open OpenVLA-UAV/vla-scripts/finetune_uav.sh and update all local paths, including the pretrained model path, dataset path, run directory, adapter directory, and Weights & Biases settings.

Then launch finetuning:

cd OpenVLA-UAV
bash vla-scripts/finetune_uav.sh

Evaluation

Download and extract the packaged UnrealZoo environment:

Collection_WinNoEditor_0424_25.zip

Configure the simulator path in:

UAV-Flow-Eval/gym_unrealcv/envs/setting/Track/DowntownWest.json

Update env_bin_win to the extracted Unreal environment executable path.

Start the OpenVLA-UAV inference server:

cd OpenVLA-UAV
python vla-scripts/openvla_act.py

In another terminal, run UAV-Flow-Eval:

cd UAV-Flow-Eval
python batch_run_act_all.py

To use a different inference server port:

python batch_run_act_all.py --server_port 5006

After evaluation finishes, compute nDTW metrics:

python metric.py

📝 Citation

If you find this work useful, please consider citing our paper:

@article{wang2026uav,
  title={Uav-flow colosseo: A real-world benchmark for flying-on-a-word uav imitation learning},
  author={Wang, Xiangyu and Yang, Donglin and Liao, Yue and Zheng, Wenhao and Dai, Bin and Li, Hongsheng and Liu, Si and others},
  journal={Advances in Neural Information Processing Systems},
  volume={38},
  year={2026}
}

📄 License

This project is licensed under the Apache-2.0 License. See LICENSE for more information.

🙏 Acknowledgement

This project builds on OpenVLA, UnrealZoo Gym. We thank the maintainers of these projects for releasing their code, models, and simulation resources.