Path Planning Strategy Optimization of Autonomous Driving Decision System Based on Reinforcement Learning

To address the challenge of balancing safety and efficiency in dynamic and constrained traffic environments, this study employs the Twin Delayed Deep Deterministic Policy Gradient (TD3) algorithm with a dual critic network to mitigate Q-value overestimation and optimize path planning in a continuous action space. A multi-dimensional state vector— including vehicle position, dynamic targets, traffic light status, and historical actions—is constructed, alongside a continuous velocity-curvature action space and a composite reward function incorporating time minimization, collision/violation penalties, and acceleration constraints. A safety barrier corrects high-risk actions in real time. Experiments on the LG Silicon Valley Lab Simulator show that under high traffic density (40 vehicles/km), TD3 reduces average driving time and path length by up to 16.8% and 3.1 km compared to Deep Deterministic Policy Gradient (DDPG) and Asynchronous Advantage Actor-Critic (A3C). TD3 also achieves significantly lower collision rates in adverse weather. These results confirm TD3's effectiveness in enhancing both safety and efficiency for autonomous driving decision-making systems.