Last week we participated in the European Workshop on Reinforcement Learning (EWRL 2026), presenting two posters showcasing our work on reinforcement learning and multi-agent systems.

🔹 URB – Urban Routing Benchmark for RL-equipped Connected Autonomous Vehicles introduced a comprehensive benchmarking environment for evaluating reinforcement learning methods for urban routing across 29 real-world traffic networks with realistic demand patterns. The benchmark provides standardized tasks, multi-agent reinforcement learning implementations, baseline methods, and domain-specific performance metrics. Experimental results reveal that state-of-the-art methods rarely outperform human routing performance despite substantial training costs, and continue to face significant scalability challenges.

🔹 PC3D: Zero-Shot Cooperation Across Variable Rosters via Personalized Context Distillation introduced a method for enabling decentralized agents to cooperate in teams of varying sizes without execution-time communication or online retraining. By distilling coordination context from a centralized teacher into decentralized policies, PC3D allows agents to infer personalized context from their local interaction histories and adapt their decision-making accordingly. Experiments across three cooperative multi-agent reinforcement learning benchmarks demonstrated higher returns than the evaluated baselines for both seen and unseen team sizes.