language:
- en
- ja
license: cc-by-nc-4.0
pretty_name: 360CityArena
arxiv: 2608.08814
size_categories:
- n<1K
tags:
- benchmark
- embodied-ai
- multimodal
- visual-reasoning
- navigation
- 360-degree
- unity
- image
- tabular
- datasets
360CityArena
360CityArena is a Unity-based embodied navigation and visual reasoning benchmark for multimodal large language models. This dataset release contains 175 tasks across seven task families and the reference images required by those tasks.
August 11, 2026: Task 7006 was corrected after we found that it duplicated task 7009.
Loading
from datasets import load_dataset
dataset = load_dataset("hal-utokyo/360CityArena")
test = dataset["test"]
print(test.num_rows) # 175
print(test[0])
The reference_image field is an image for tasks that require one and None
otherwise. There are 75 task rows with a reference image and 51 unique image
files: one localization map shared by 25 tasks, 25 landmark images, and 25
navigation maps.
Task inventory
| Task family | Tasks |
|---|---|
| Localization | 25 |
| Landmark Search with Language | 25 |
| Landmark Search with Image | 25 |
| Counting | 25 |
| Map Navigation | 25 |
| Language Guided Navigation | 25 |
| Relational Spatial Reasoning | 25 |
All records are in the test split because 360CityArena is an evaluation
benchmark, not a training corpus.
Record fields
Each row contains the task ID and family, difficulty, initial environment index,
task prompt, ground-truth answer, and task-specific goal fields.
reference_image_path identifies the corresponding runner asset while
reference_image provides the image directly through 🤗 Datasets and the Dataset
Viewer. The legacy source_file, source_row, and source_record fields retain
provenance from the initial Hub migration.
Intended use and limitations
The dataset is intended for evaluating embodied multimodal agents with the 360CityArena Unity environment and runner. Loading this dataset alone does not run the interactive benchmark; use the code and evaluation protocol in the GitHub repository. The benchmark covers a fixed virtual reconstruction of an urban area and should not be treated as evidence of general navigation ability, physical-world safety, or geographic coverage. Some reference images contain Japanese storefront text and other scene-specific visual information.
License and attribution
Dataset records and benchmark assets are distributed under the Creative Commons
Attribution-NonCommercial 4.0 International license (CC BY-NC 4.0), except where
third-party terms apply. Map-derived assets include OpenStreetMap data:
copyright OpenStreetMap contributors, available under the Open Database License
(ODbL). See the repository's DATA_LICENSE and NOTICE files for details.
The benchmark runner and other source code are separately licensed under Apache-2.0 in the GitHub repository.
Citation
@inproceedings{watanabe2026360cityarena,
title = {360CityArena: A Realistic Virtual Urban Navigation Benchmark for Embodied Agents},
author = {Watanabe, Kenta and Miyai, Atsuyuki and Takenawa, Mizuki and Aizawa, Kiyoharu and Yamasaki, Toshihiko},
booktitle = {Proceedings of the European Conference on Computer Vision (ECCV)},
year = {2026}
}