Safe control of autonomous vehicles in overtaking maneuvers using game-theoretic learning-based predictive controller

This work proposes a safe control strategy for an autonomous vehicle to overtake a human-driven vehicle (HDV) using a predictive safety filter (PSF) mechanism that hierarchically combines an end-to-end Reinforcement Learning (RL) agent with a predictive controller. To create a more realistic RL environment, a Stackelberg game based on a first principles model is employed to capture the HDV’s real-time response during overtaking rather than relying on a predefined empirical or purely statistical driver model. In the lower layer, a distributionally robust chance-constrained predictive controller is implemented to manage uncertainties in HDV behavior, ensuring robust safety guarantees. The effectiveness of the proposed synthetic controller is verified in a gym environment with comparisons against traditional schemes.