Safety testing and enhancement of urban traffic signals for safe deployment

Urban deployments of AI-driven, adaptive traffic signal control (TSC) systems are advancing rapidly, yet systematic pre-deployment safety validation of these "black-box" models remains underexplored. Without rigorous safety validation, these models may pose potential risks when deployed at real-world intersections. We present a model-agnostic framework that adversarially tests any given TSC model within a virtual traffic environment, using deep reinforcement learning to expose their safety-critical "failure points" in travel demand distributions. The expected safety risk of each given model is quantified through a metric that integrates surrogate crash severity with the empirical likelihood of hazardous demand patterns, grounded in over 11 years of real-world traffic signal volume data. Building on these insights, we propose a safety-critical re-training (SCRT) mechanism for safety-underperforming models, which selectively enriches training with the uncovered "failure point" to enhance robustness without altering the model architecture. Across diverse scenarios, this approach consistently reduces expected safety risk following SCRT and significantly outperforms naive re-training that disregards safety-critical demand patterns. Importantly, these safety gains are achieved with only slight to modest impacts on operational efficiency. By operationalizing a "test-before-deploy" safety gate, our framework provides a practical pathway from design-time optimization to deployment-time safety assurance, ultimately advancing safer real-world adoption of AI-driven TSC systems.