Last-Meter Precision Navigation for UAVs: A Diffusion-Refined Aerial Visual Servoing Approach

---

From coarse aerial localization to fine-grained last-meter navigation: a diffusion-refined visual servoing framework for precise UAV landing and delivery.

Project Structure

UAVM_2026/
├── models/
│   ├── dino_resnet/
│   └── controlnet/
├── pairUAV/
│   ├── data_process.sh
│   └── University-Release.zip
├── baseline/
│   ├── SuperGlue/
│   ├── train.py
│   └── run.sh
└── step2_refine/
    ├── train_rgb_loss.py
    ├── train_rgb_condition_predictor.py
    ├── tutorial_dataset.py
    ├── cldm/
    ├── ldm/
    ├── cldm_v15_pose_hybrid.yaml
    ├── train_step2_example.sh
    └── train_rgb_condition_predictor_example.sh

---

1. Environment Setup

Create a unified conda environment for the baseline:

conda create -n uavm python=3.9
conda activate uavm

pip install torch==2.7.1 torchvision==0.22.1 torchaudio==2.7.1 --index-url https://download.pytorch.org/whl/cu128
pip install -r requirements.txt

huggingface-cli download Ramos-Ramos/dino-resnet-50 --local-dir models/dino_resnet

For Stage-II diffusion training, additional dependencies from latent diffusion / ControlNet may be required. Please install the dependencies listed in the Step-II environment file if provided.

---

2. Data Preparation

2.1 Download University-1652 Dataset

Download University-1652 upon request. You may use the request template.

2.2 Download and Process PairUAV Dataset

Download and process the PairUAV dataset:

cd pairUAV/
bash data_process.sh
cd ..

This script downloads the dataset from HuggingFace and extracts train/test/tours data to the pairUAV/ directory.

---

3. Stage-I: SuperGlue-Based Coarse Pose Estimation

3.1 Run SuperGlue Feature Matching

First, perform feature matching on image pairs:

cd baseline/SuperGlue

# Option 1: download precomputed matching results.
bash download_results.sh
cd ..

# Option 2: run feature matching.
python gen_test_pairs.py
bash run_train.sh
bash run_test.sh
cd ..

This generates matching results in train_matches_data/ and test_matches_data/.

3.2 Train Stage-I Model

cd baseline/
bash run.sh
cd ..

The Stage-I model predicts coarse heading and range from an image pair. The predicted pose can be exported as a JSON file and used as the condition input for Stage-II diffusion refinement.

3.3 Evaluate Stage-I Results

The final evaluation is conducted on CodaBench. After generating your test predictions, package the submission files according to the competition requirements and upload them to:

https://www.codabench.org/competitions/15251/

Note:

• The official test results are only available through the CodaBench evaluation server.
• Please make sure your submission file strictly follows the format required by the competition page.
• Local validation can be used for debugging, but the leaderboard scores on CodaBench are the final results used for comparison.

---

4. Stage-II: Diffusion-Based Next Observation Generation

Stage-II trains a diffusion refinement model for next-observation generation. The model is a ControlNet-style latent diffusion model conditioned on:

• a source RGB image through the ControlNet hint pathway;
• a numeric pose condition, including heading and range, through a trainable pose encoder;
• optionally, a frozen RGB pose predictor used as an auxiliary pose-consistency loss.

4.1 Main Files

step2_refine/
├── train_rgb_loss.py
├── train_rgb_condition_predictor.py
├── tutorial_dataset.py
├── cldm/
├── ldm/
├── cldm_v15_pose_hybrid.yaml
├── train_step2_example.sh
└── train_rgb_condition_predictor_example.sh

• train_rgb_loss.py: main Stage-II diffusion training script.
• train_rgb_condition_predictor.py: trains the frozen RGB pose predictor used by the auxiliary RGB pose-consistency loss.
• tutorial_dataset.py: PairUAV dataset loader for Stage-II training.
• cldm/: ControlNet and DreamNav model components.
• ldm/: latent diffusion model components.
• cldm_v15_pose_hybrid.yaml: Stage-II model configuration.
• train_step2_example.sh: example Stage-II training script.
• train_rgb_condition_predictor_example.sh: example RGB pose predictor training script.

4.2 External Files Required

The following files are not included in this repository:

• PairUAV dataset;
• base ControlNet checkpoint, e.g. control_sd15_ini.ckpt;
• Stage-I pose JSON, e.g. step1_train_truepose.json or predicted pose JSON;
• frozen RGB pose predictor checkpoint, e.g. best.pt.

4.3 Train RGB Pose Predictor

The RGB pose predictor takes a source RGB image and a target/generated RGB image as a 6-channel input pair, and predicts:

[sin(heading), cos(heading), range / range_scale]

Train it with:

cd step2_refine/
bash train_rgb_condition_predictor_example.sh

The generated best.pt can be used as a frozen auxiliary model during Stage-II diffusion training.

4.4 Train Stage-II Diffusion Model

After preparing the dataset, base checkpoint, Stage-I pose JSON, and optional RGB predictor checkpoint, run:

cd step2_refine/
bash train_step2_example.sh

Please edit dataset paths, checkpoint paths, and output paths in the bash scripts before running.

4.5 Default Trainable Scope

For the default train_mode=lora_control_decoder_hint setting:

• the VAE is frozen;
• most pretrained diffusion backbone weights are frozen;
• the pose encoder is trainable;
• LoRA adapters are trained in selected ControlNet and UNet decoder linear layers;
• the ControlNet hint pathway is trainable;
• the RGB pose predictor is frozen and used only for auxiliary pose-consistency loss.

---

5. Files Not Included

Do not commit large datasets, checkpoints, or generated outputs:

*.ckpt
*.pt
*.pth
*.safetensors
outputs*/
checkpoints/
pairUAV/
matches_data/
train_matches_data/
test_matches_data/
step1_*.json
lightning_logs/
wandb/
__pycache__/
*.pyc

🔗 Ecosystem

Explore our ecosystem for UAV & Spatial Intelligence 🚁

🚁 UAV & Spatial Intelligence

🎓 The University-1652 Family

🎓 University-1652 Multi-view Multi-source Benchmark Ground · Drone · Satellite · ACM MM'20

🌦️ University-WX Multi-Weather Extension on the Fly Pattern Recognition'24

💬 GeoText-1652 Dense Text Extension ECCV'24

🚀 New Open-Source Releases

🛰️ GeoFuse Road Maps as Free Geometric Priors

🧠 UAVReason Aerial Scene Reasoning & Generation Benchmark

🗺️ Video2BEV Drone Video → Bird's-Eye-View

🚁 PairUAV Paired UAV Data for Matching

---

⭐ If you find our projects helpful, a star is the best support! ⭐

📖 Citation

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

arXiv Preprint

@misc{li2026lastmeterprecisionnavigationuavs,
      title={Last-Meter Precision Navigation for UAVs: A Diffusion-Refined Aerial Visual Servoing Approach}, 
      author={Yaxuan Li and Jiarui Zeng and Shaofei Huang and Zhedong Zheng},
      year={2026},
      eprint={2607.04352},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2607.04352}, 
}

ACM Multimedia Workshop

@inproceedings{deuser2026UVA,
  title={The 4th Workshop on UAVs in Multimedia: Capturing the World from a New Perspective},
  author={Deuser, Fabian and Li, Yaxuan and Wang, Tingyu and Shi, Yujiao and Bößendörfer, Anna and Huang, Shaofei and Pan, Xiao and Zheng, Zhedong and Zimmermann, Roger},
  booktitle={Proceedings of the 34th ACM International Conference on Multimedia Workshop},
  year={2026}
}