Adaptive traffic signal optimization with thermal sensors and reinforcement learning

• Real-world adaptive traffic signals with thermal sensors cut vehicle delays significantly • A new learning control system adjusts timing using traffic and environmental conditions • Safety guard rules ensure every signal change follows required road rules at all times • The learning system reduces fuel use and carbon emissions while improving travel times • Uses field data to build a realistic model that prepares the system for city deployment Static traffic signals remain a primary source of urban congestion, increasing vehicle delays and environmental impact. While reinforcement learning (RL) offers a promising solution, existing models often fail to address the multi-objective nature of traffic management, specifically by overlooking environmental costs and neglecting the operational safety guarantees required for real-world deployment. This study addresses this gap by introducing a novel, safety-first adaptive control framework, validated through a two-stage methodology. Initially, a rule-based adaptive system using thermal sensors was deployed in Dnipro, Ukraine, establishing a real-world performance baseline with a 20–38 % delay reduction over fixed-time plans. Subsequently, a novel multi-agent RL (MARL) framework, Dynamic, Cost-Aware MARL (DCA-MARL), was developed and calibrated against this field data. The framework’s core innovations include a dynamic, multi-objective reward function that balances traffic efficiency with emissions, and a hard-coded Safety Constraint Layer (SCL) to ensure 100 % compliance with operational mandates. High-fidelity microscopic simulations demonstrate that DCA-MARL significantly outperforms fixed-time, the real-world actuated baseline, and standard RL models, achieving substantial reductions in delay while simultaneously and dramatically cutting fuel consumption and CO 2 emissions, all while maintaining zero safety violations.