Real topology
Networks from รle-de-France and Ingolstadt varying in size and complexity.
NeurIPS 2025
Urban Routing Benchmark
for RL-equipped Connected Autonomous Vehicles
A comprehensive benchmarking environment unifying evaluation across 29 real-world traffic networks to test multi-agent reinforcement-learning routing strategies in mixed-autonomy traffic.
Contributors
The research team
Ahmet Onur AkmanJagiellonian University
Anastasia PsarouJagiellonian University
Michaล HoffmannJagiellonian University
ลukasz GorczycaJagiellonian University
ลukasz KowalskiUrban Policy Observatory
Paweล GoraJagiellonian University
Grzegorz JamrรณzJagiellonian University
Rafaล KucharskiJagiellonian University
The future of urban routing
Connected Autonomous Vehicles (CAVs) promise to reduce congestion by optimizing routing decisions collectively. However, standard benchmarks for this complex multi-agent problem have been missing.
URB fills this gap. It integrates the microscopic traffic simulator SUMO with RL-based solutions to create a realistic testing ground. It features 29 real-world networks, calibrated demand patterns, and a suite of baseline algorithms.
Realistic demand
Calibrated trip patterns based on census and empirical traffic data.
Standard metrics
Evaluate travel time, congestion, and network-level efficiency.
Open data
Fully reproducible datasets and modular configuration schemes.
The platform
From map to metric
URB standardizes the pipeline for setting up traffic scenarios, training MARL agents, and assessing their impact on the city.
Setup
29 real-world networks and demand patterns
URB dataset โTraining
MARL algorithms including IQL and QMIX
Assessment
Compare travel times and congestion
URB metricsDiverse environments
URB includes networks ranging from small towns such as St. Arnoult to dense urban centers such as Ingolstadt.
- 1St. Arnoult
Small scale, 222 trips. Good for debugging and quick iteration.
- 2Provins
Medium scale, 523 trips. Represents typical suburban traffic.
- 3Ingolstadt
Large scale, 1,035 trips. High congestion and challenging coordination.
Network atlas
SUMO ร RouteRL
01
02
03
Microscopic simulation
Powered by SUMO, URB simulates individual vehicle dynamics to provide realistic feedback to learning agents, going beyond fluid approximation models.
Dynamics
Car-following physics and gap acceptance
Intersections
Traffic lights, right-of-way rules, and junctions
Mixed traffic
Interaction between human drivers and algorithmic CAV agents
Benchmark results
We evaluated state-of-the-art MARL algorithms against URB baselines. The results reveal a significant challenge: current algorithms rarely outperform human drivers in complex, congested networks.
Results reported in URB (arXiv:2505.17734) โSt. Arnoult (Small)
Mean CAV travel times. Lower is better. In small networks, QMIX occasionally beats humans.
Baselines
Human drivers3.15m
All-or-nothing3.01m
Random3.58m
MARL algorithms
QMIX3.21m
IPPO3.33m
IQL3.53m
MAPPO3.51m
In this scenario, RL agents achieved a win rate of 80% against URB baselines.
Scaling issues
Algorithms that work on small maps fail to converge or improve upon baselines in larger networks.
Cost of training
Exploring suboptimal routes can cause regret and delays for commuters during the learning phase.
Future work
URB tracks progress toward methods that handle non-stationarity and massive scale.
Quick start
Start benchmarking
Clone the repository, install the dependencies, and run a first experiment.
- # 1. Clone the repository
$ git clone https://github.com/COeXISTENCE-PROJECT/URB.git - # 2. Install dependencies
$ cd URB && pip install -r requirements.txt - # 3. Run an example experiment
$ python scripts/qmix_torchrl.py --id demo --alg-conf config3 --task-conf config4 --net saint_arnoult
Join the leaderboard
Implement a solution using the provided template, run it like any other URB experiment, and submit the results for comparison.
How to contribute
Use the template script and documentation, then submit the resulting experiment through a pull request.
Citation
@inproceedings{URB,
title={URB -- Urban Routing Benchmark for RL-equipped Connected Autonomous Vehicles},
author={Akman, Ahmet Onur and Psarou, Anastasia and Hoffmann, Michaล and Gorczyca, ลukasz and Kowalski, ลukasz and Gora, Paweล and Jamrรณz, Grzegorz and Kucharski, Rafaล},
booktitle={Proceedings of the Thirty-Ninth Annual Conference on Neural Information Processing Systems (NeurIPS 2025) Datasets and Benchmarks Track},
month={December},
year={2025}
}
Supported by
