RobustLight++: A Meta-Diffusion Framework for Robust Traffic Signal

Mingyuan Li, Chunyu Liu, Zhuojun Li, Xiao Liu, Guangsheng Yu, Qianrun Chen, Lixiang Li, Haipeng Peng

Keywords: diffusion model, meta learning

Abstract: Despite remarkable progress in Reinforcement Learning (RL) for Traffic Signal Control (TSC), existing methods largely lack the ability to generalize across cities, limiting their applicability in real deployments. The recent SoTA method RobustLight improves robustness but still exhibits weak transfer performance, high inference latency, and limited resistance to sensor failures. In this paper, we present RobustLight++, a meta-diffusion-based framework designed to explicitly learn transferable representations among heterogeneous urban environments. By theoretically linking DDIM with Reptile meta-learning, RobustLight++ enables a diffusion policy that supports both zero-shot deployment and few-shot adaptation in unseen cities, significantly reducing the cost of retraining and data collection in new domains. Comprehensive experiments on large-scale real-world benchmarks demonstrate superior cross-city transfer capability, with performance gains ranging from 7.41\% to 52.13\% under diverse noise conditions, and consistent improvements over all competing baselines in unseen environments. In addition, RobustLight++ achieves up to 91.9\% reduction in inference latency, ensuring real-time applicability. The proposed framework delivers a practical solution toward scalable, transferable, and robust urban traffic control systems. Our code is available at https://anonymous.4open.science/r/RobustLightPlus-E14F.

Primary Area: other topics in machine learning (i.e., none of the above)

Submission Number: 6454