Curriculum Reinforcement Learning for Robust Autonomous Vehicle Entry in Mixed-Autonomy Roundabouts

Autonomous vehicle (AV) insertion into mixed-autonomy roundabouts is a challenging decision-making problem due to interactions with human-driven vehicles (HDVs), uncertain traffic gaps, and collision risks. This work investigates a curriculum reinforcement learning framework for training a Proximal Policy Optimization (PPO) agent to perform robust AV insertion in a simulated roundabout environment using Simulation of Urban MObility (SUMO). The framework combines a spatial spawn curriculum with a traffic-composition curriculum that progressively increases the proportion of human-driven vehicles during training. The approach is evaluated under different HDV penetration rates, with performance assessed using insertion success, collision rate, timeouts, and safety-related metrics. The study also examines training behavior and selected ablation configurations to investigate the effects of the proposed training strategy. The results are specific to the simulation environment and evaluation settings; further validation is required to establish generalization to real-world traffic. This manuscript presents simulation-based research and has not undergone peer review.