Safety-Constrained Reinforcement Learning Method for Urban Eco-Driving

Reinforcement Learning (RL) is used for eco-driving in Intelligent and Connected Vehicles (ICVs), but traditional methods face a critical trade-off. Ensuring safety often requires high penalty terms, leading to overly conservative policies that severely compromise energy and traffic efficiency. To resolve this conflict, this paper introduces the Twin Delayed Deep Deterministic Policy Gradient with Safelayer (TD3+Safelayer), a novel framework that decouples safety from performance objectives. The Safelayer module maps the TD3 policy's actions into a verifiably safe action space. SUMO simulations demonstrate that TD3+Safelayer eliminates all collision events present in the baseline models. While achieving this high level of safety, it also reduces energy consumption by 17.4% and 30.7% compared to traditional TD3 and rule-based strategies, respectively.