AerialDojo-200K is a large-scale benchmark suite for open-world aerial object-goal search. Please provide your affiliation and intended use to request access.

Please provide accurate information and review the dataset license and documentation before submitting your request.

AerialDojo-200K

AerialDojo-200K is a large-scale benchmark suite for open-world aerial object-goal search. Aerial agents autonomously explore large-scale, unstructured 3D environments and reach target objects specified by semantic descriptions (SemanticOGS) or reference images (ImageOGS), rather than follow route-specific instructions.

Paper: arXiv · Project: Website · Code and setup: GitHub

Large-Scale. 42 simulation scenes span four scene families and 21 scene types, with 205,732 task instances across SemanticOGS and ImageOGS under Base, Standard, and Long-Horizon settings. The suite offers 3× as many scenes and 18.7× as many task instances as the largest prior benchmark for aerial object-goal search.
High-Quality. 12 annotators spent two months annotating 109 landmarks, 2,099 target objects, and 2,099 object anchors. Collision-free reference trajectories cover 4,115.313 km of unique routes, complemented by 63,177 groups of multi-view recordings collected exclusively for training.
Unified Evaluation. Shared data formats, action spaces, and evaluation protocols support evaluation across 21 in-distribution and 21 out-of-distribution scenes. Baseline evaluations of nine multimodal large language models—five open-source and four closed-source—highlight the remaining challenges in building general-purpose aerial agents.

OOD test evaluation

The out-of-distribution test split (OOD_TESTS) is not publicly released. To evaluate your method on this split, please contact the authors through Email to arrange testing.

The public task and trajectory splits are ID_TRAINS, ID_TESTS, and OOD_TRAINS. OOD test tasks, reference images, and trajectories are withheld. Both ID_ENVS and OOD_ENVS map environments are available.

Dataset contents

Directory	Contents
`AerialENVS/`	Packaged UE / ProjectAirSim environments
`SemanticOGS/`	Semantic-goal search tasks in `Task.json`
`ImageOGS/`	Image-goal search tasks and reference images
`TrajectoryDATA/`	One merged `Trajectories.json` per map/task/split partition

Within the same partition, a task's episode_id = "n" corresponds to trajectories["n"] in Trajectories.json. SemanticOGS and ImageOGS provide two goal representations of the same object-goal search tasks.

Usage

Clone the code repository, then run the following from its root to download the data into the matching directories:

hf download fengtt42/AerialDojo-200K --repo-type dataset --local-dir . \
  --include 'AerialENVS/*' \
  --include 'SemanticOGS/*' \
  --include 'ImageOGS/*' \
  --include 'TrajectoryDATA/*'

The hf command is provided by huggingface_hub. Follow the GitHub README for installation, trajectory recording, and running your own policy.

Citation

If you use AerialDojo-200K in your research, please cite our paper:

@article{feng2026aerialdojo,
  title={AerialDojo-200K: A Large-Scale Benchmark Suite for Open-World Aerial Object-Goal Search},
  author={Feng, Tongtong and Wang, Xin and Hou, Haoran and Wang, Ren and Wang, Weiran and Zhu, Shaokai and Jia, Ziqi and Wang, Hao and Zhan, Yu-Wei and Wu, Zongyuan and Cui, Jinghao and Zhu, Wenwu},
  journal={arXiv preprint arXiv:2609.36066},
  year={2026}
}

Downloads last month: 86

Total file size:

130 GB

Paper for fengtt42/AerialDojo-200K

AerialDojo-200K: A Large-Scale Benchmark Suite for Open-World Aerial Object-Goal Search

Paper • 2609.36066 • Published 4 days ago