README.md

metadata

license: apache-2.0
pretty_name: VehicleArena Road Networks

VehicleArena Road Networks

Offline map assets for VehicleArena's 3D urban-driving benchmark

English project README · 中文项目说明

This repository contains the versioned road-network bundle used by VehicleArena. The simulator, benchmark scenarios, evaluation code, and map installer live in the GitHub repository; the large offline maps are distributed here.

Start here: Download · Install · Benchmark context · Citation

🌟 Highlights

• 116 road-network maps: the archive contains 116 base road networks and their 116 lane-level counterparts, for 232 map JSON files in total.
• Versioned and verifiable: the archive follows CalVer YYYY.MM.DD.PATCH; the VehicleArena installer checks its SHA-256 and the individual map files against the code repository's manifest.
• Maps, not trajectories: task definitions and evaluation software are maintained on GitHub, so the same map bundle can be used with the published tasks or new scenarios.

📦 Download

Item	Value
Version	2026.09.22.0
Archive	vehiclearena-road-networks-2026.09.22.0.tar.gz
Contents	116 base + 116 lane-level networks (232 JSON files)
Size	117,275,038 bytes
SHA-256	79e9b642db23c37373697b5431dd615846c57b7c9b7e2c520c64d4ed3d608f75

The archive can also be downloaded with the Hugging Face CLI:

hf download OpenMOSS-Team/VehicleArena \
  vehiclearena-road-networks-2026.09.22.0.tar.gz \
  --repo-type dataset --local-dir /path/to/downloads

🚀 Installation

First set up the VehicleArena code and dependencies. From the code repository root, install the downloaded archive and check the environment:

python scripts/manage_map_bundle.py install \
  --source /path/to/downloads/vehiclearena-road-networks-2026.09.22.0.tar.gz
python scripts/check_environment.py

The check reports "ready": true when Python, SUMO, and the maps are available. See the map installation guide for the archive layout and checksum validation.

📊 Benchmark context

The published VehicleArena benchmark has 100 training/development tasks and 112 held-out evaluation tasks: 80 Basic and 32 MultiLLM. Those tasks, their manifests, and the evaluation code are in GitHub, not in this map archive. The published time limits were calibrated using this map version; changing road data or task definitions requires recalibrating affected scenarios. Reusing the same archive with another evaluated model does not.

📎 Citation

If you use these maps with VehicleArena, please cite the paper:

@article{yang2026vehiclearena,
  title={VehicleArena: A Realistic Urban Environment for Multi-Agent Driving},
  author={Yang, Jie and Chen, Jiajun and Zhou, Jiazheng and Huang, Mianqiu and Zheng, Yining and Wang, Yuxin and Qiu, Xipeng},
  journal={arXiv preprint arXiv:2609.35916},
  year={2026}
}