Research map / Multi-agent systems and games / Human-AI and goal-conditioned
URB - Urban Routing Benchmark for RL-equipped Connected Autonomous Vehicles
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Introduces URB, a benchmark of 29 real-world traffic networks in SUMO (via RouteRL) for MARL-controlled connected autonomous vehicle routing, and benchmarks IQL, IPPO, MAPPO, and QMIX against random, all-or-nothing, and human baselines.
Problem
Standardized, realistic benchmarks are missing for evaluating reinforcement learning routing strategies for connected autonomous vehicles sharing urban road networks with human drivers.
Approach
A modular benchmark that couples MARL with a microscopic SUMO simulator and realistic Île-de-France and Ingolstadt demand, with a human day-to-day route-choice model, mutation from human to CAV agents, and metrics such as training cost and winrate against human travel times.
Key results
- Only QMIX on St. Arnoult outperformed human pre-CAV travel times (80% winrate), and it performed worse than random in the more congested Provins and Ingolstadt networks.
- IQL and MAPPO failed to converge and reached performance nearly at random across networks.
- In the sensitivity analysis, IQL improvement over pre-CAV times fell from -5.5% at half demand to -79.3% at double demand.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use QMIX instead of human-equivalent routing | CAV routing · St. Arnoult | CAV mean travel time: 3.15 -> 3.21 | Helped |
| Use QMIX instead of random routing | CAV routing · St. Arnoult | CAV mean travel time: 3.58 -> 3.21 | Helped |
| Use QMIX instead of random routing | CAV routing · Provins | CAV mean travel time: 3.04 -> 3.14 | Hurt |
| Use MARL routing instead of human-only routing | St. Arnoult · St. Arnoult | mean travel time: 3.15 -> 3.24 | Mixed |
| Increase demand from half to original (ablation) | IQL · St. Arnoult | CAV travel time improvement rate: -5.50% -> -12.06% | Hurt |
| Increase demand from original to double (ablation) | IQL · St. Arnoult | CAV travel time improvement rate: -12.06% -> -79.32% | Hurt |
| Use global route-selection observations instead of local observations (ablation) | IQL · St. Arnoult | CAV travel time improvement rate: -12.06% -> -3.49% | Helped |
| Increase demand from half to original (ablation) | MAPPO · St. Arnoult | CAV travel time improvement rate: -5.20% -> -9.52% | Hurt |
| Increase demand from original to double (ablation) | MAPPO · St. Arnoult | CAV travel time improvement rate: -9.52% -> -61.42% | Hurt |
| Train QMIX for longer (ablation) | QMIX · Scenario 1, two instances | performance | Mixed |
Surprising findings
- Community implementations of state-of-the-art MARL algorithms rarely outperformed humans or even random routing.
- Doubling demand causes sharp degradation, and human travel times also increase after CAV deployment.
Limitations
- Same driving model for humans and CAVs; CAV traffic-flow benefits are neglected.
- Routes are chosen once per episode and demand is fixed in OD pairs and departure times.
- Only a single CAV fleet is considered and the experimental scheme is limited in algorithms and scenarios.
Experimental setup
Datasets: Île-de-France synthetic population and travel demand (28 subregions) InTAS Ingolstadt scenario from RESCO
Benchmarks: URB Scenario 1 (40% CAVs) on St. Arnoult, Provins, Ingolstadt URB sensitivity analysis on St. Arnoult
Metrics: mean travel time training cost mean speed change mileage change winrate
Baselines: IQL IPPO MAPPO QMIX All-or-Nothing Random Human
Builds on: RouteRL TorchRL SUMO QMIX MAPPO
Closest papers by meaning
- One Step is Enough: Multi-Agent Reinforcement Learning Based on One-Step Policy Optimization for Order Dispatch on Ride-Sharing Platforms (NeurIPS 2026 · similarity 0.87)
- IntersectionZoo: Eco-driving for Benchmarking Multi-Agent Contextual Reinforcement Learning (ICLR 2025 · similarity 0.87)
- Beyond Self-Play and Scale: A Behavior Benchmark for Generalization in Autonomous Driving (NeurIPS 2026 · similarity 0.86)
- Advancing Multi-agent Traffic Simulation via R1-Style Reinforcement Fine-Tuning (ICLR 2026 · similarity 0.85)
- PARCO: Parallel AutoRegressive Models for Multi-Agent Combinatorial Optimization (NeurIPS 2025 · similarity 0.85)
- Learning to Drive in New Cities Without Human Demonstrations (NeurIPS 2026 · similarity 0.84)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.