Optimizing roundabout management via deep reinforcement learning with safety and comfort constraints

This paper presents a deep reinforcement learning (DRL) framework to optimize autonomous vehicle maneuver during roundabout approaches, with a focus on safety, efficiency, and passenger comfort. The proposed method incorporates a logistic regression-based Roundabout Exit Probability (REP) model to estimate the likelihood that inbound vehicles will exit the roundabout, as well as a regression-based Time-To-Collision (TTC) predictor to model the ego vehicle's controlled maneuver while maintaining comfort constraints. These predictive models are integrated into a Proximal Policy Optimization (PPO) framework, enhanced with a curriculum learning strategy to gradually shape the agent's behavior toward balanced, human-like decisionmaking. The reward function is designed to penalize unsafe or abrupt actions and encourage smooth, efficient maneuvering. Experimental results in the CARLA simulator demonstrate the effectiveness of the proposed strategy in achieving robust, comfort-aware navigation in roundabout scenarios.