AerointentBenchmark/AeroIntent-city

AeroIntent: City Scene

This repository contains the City environment of AeroIntent, a synchronized multimodal benchmark for UAV mission intent recognition.

The dataset is now publicly available for non-commercial use. Downloads are open to everyone, with no access request or manual approval required.

Paper: AeroIntent: A Synchronized Multimodal Benchmark for UAV Mission Intent Recognition, arXiv:2610.25869, 2026.

Please cite our paper when using AeroIntent in your research or publications. Commercial use is not permitted under the dataset license.

Benchmark Repositories

The benchmark comprises 990 synchronized sequences across three environments:

Environment	Dataset
City	AeroIntent-city
Landscape	AeroIntent-landscape
Park	AeroIntent-park

Contents

• 15 UAV configurations from three airframe families and five payload configurations
• 22 trajectory programs grouped into Transit, Reconnaissance, Probing, Strike, and Escape
• RGB, infrared, depth, Echo radar, IMU, GPS, bounding boxes, timestamps, and metadata
• 330 synchronized sequences in this environment

User Guide

See the AeroIntent Dataset User Guide for the sequence layout, modality files, field meanings and units, timestamps, IMU/GPS, radar returns, bounding boxes, and key metadata.

Download and Reconstruction

The data are stored as one tar.zst shard per UAV configuration. Each shard is split into ordered 256-MiB parts. Download the parts and manifest for a UAV configuration, reconstruct and verify its shard, then extract all shards into the same destination directory.

For example, using the Hugging Face CLI:

hf download AerointentBenchmark/AeroIntent-city --repo-type dataset --local-dir AeroIntent-city
cd AeroIntent-city
python reconstruct_shard.py data/Drone_no_payload
tar --use-compress-program=unzstd -xf Drone_no_payload.tar.zst

Each UAV directory contains a manifest.json with the original archive checksum and ordered part checksums. SHA256SUMS.txt lists all uploaded parts. Collection logs, batch logs, and completion markers are excluded. The released sequences use the view_front directory shown below.

Dataset Structure

<UAV configuration>/<intent>/<trajectory>/view_front/
  rgb/ ir/ depth/ depth_preview/
  echo/ echo_groundtruth/ imu/ gps/
  labels_all.txt timestamps.csv meta.json
  rgb_video.mp4 ir_video.mp4 depth_video.mp4

Citation

If you use AeroIntent, please cite:

@misc{chen2026aerointent,
  title         = {{AeroIntent}: A Synchronized Multimodal Benchmark for {UAV} Mission Intent Recognition},
  author        = {Yutong Chen and Jingxin Mao and Zhiwei Wei and Bing Li and Jianzhong Ni and Rongqing Zhang},
  year          = {2026},
  eprint        = {2610.25869},
  archivePrefix = {arXiv},
  primaryClass  = {cs.RO},
  url           = {https://arxiv.org/abs/2610.25869}
}

License and Use

The AeroIntent dataset is released under the Creative Commons Attribution-NonCommercial 4.0 International license (CC BY-NC 4.0). See LICENSE for the full terms.

• Non-commercial use only. Commercial use is not permitted under this license.
• Give appropriate credit to the AeroIntent authors and cite the paper in research or publications using the dataset.
• When sharing the dataset or adaptations, retain attribution, link to the license, and indicate any changes.

This license applies to the released dataset. Third-party software and assets remain subject to their respective licenses.