A Hybrid DDPG+MPC Framework for Safe and Efficient Autonomous Lane-Changing in Highway Overtaking

Abstract

Lane-changing decision-making is a critical component of autonomous driving, as it requires balancing safety, efficiency, and manoeuvre stability under dynamic traffic conditions. This study proposes a hybrid Deep Deterministic Policy Gradient (DDPG) framework with an MPC-inspired predictive safety layer for autonomous lane-changing in a controlled highway overtaking scenario. The DDPG policy generates candidate longitudinal commands and lateral lane-change intentions, while the supervisory layer evaluates the predicted evolution of the target-lane front gap, rear gap, and time-to-collision (TTC) over a short prediction horizon before permitting the lateral manoeuvre. Rather than solving an online MPC optimisation problem, the supervisory layer employs short-horizon state prediction and constraint-based safety assessment to determine whether the candidate lane-change intention meets the predefined safety and overtaking-necessity conditions. The proposed framework is evaluated in a unified Simulation of Urban MObility (SUMO) highway environment and compared with rule-based, MPC-only, and DDPG-only controllers using consistent scenario conditions and performance metrics. The evaluation considers task success, collision occurrence, overtaking time, average speed, safety-related spacing, driving comfort, and lane-change behaviour. The results show that all evaluated controllers completed the overtaking task without collisions under the considered scenario. However, the proposed hybrid controller achieved the shortest mean overtaking time, the highest mean speed, the largest minimum front-gap margin, and a single lane change per episode. These findings indicate that combining learning-based decision-making with lightweight short-horizon predictive safety supervision can improve overtaking efficiency and lane-change consistency while maintaining safe vehicle interactions.