VoLN-UAV Dataset

This release contains the navigation trajectories and benchmark metadata used by VoLN-UAV. Simulator environments are distributed separately.

Hugging Face Entries

• env: https://huggingface.co/datasets/Louj/VoLN-UAV-ENV
• dataset: https://huggingface.co/datasets/Louj/VoLN-UAV-dataset

Data Organization

The release contains 2,190 source-route candidates from four environments. The canonical benchmark keeps 1,786 episodes after start/goal candidate deduplication. Source candidates are retained so that the released observations and benchmark construction remain independently inspectable.

The package provides:

• scene-level Train/Validation/Test split manifests;
• route JSON files with RGB frame references and pose-derived state fields;
• episode-level active beacons, benchmark records, templates, and checksums;
• complete RGB observations under source/frames/ in the full release.

Usage

1. Download the dataset package and the env package.
2. Unzip the dataset package.
3. Set source_root in the benchmark config to the unzipped source/ directory.
4. Run python -m voln_uav.cli.build_benchmark --config <config.yaml> if the benchmark needs to be regenerated.

The generated manifest.json contains the release summary and Hugging Face resource links.

Citation

If you find this dataset useful, please consider citing the VoLN paper.

Download Layout

The dataset is uploaded as independent ZIP shards under metadata/, train/, val/, and test/. Each shard is below 5 GB. Extract metadata/VoLN-UAV-metadata.zip first, then extract the split shards you need into the same directory so that paths such as source/frames/<scene>/<trajectory>/<frame>.png match the JSONL metadata.

The train, validation, and test splits are episode-disjoint. The test split is held out on a separate scene, while train and validation may share scenes.

Use SHA256SUMS.txt to verify downloaded shards.