抄録
This study presents an integrated navigation framework for autonomous vehicles that combines Electric Potential Field (EPF), Reinforcement Learning (RL), and Model Predictive Control (MPC). The perception layer employs an EPF formulation based on finite line charges to represent obstacle bounding boxes and continuously quantify environmental risk. For decision-making, a Soft Actor-Critic (SAC) agent processes the potential-field information together with the vehicle states to determine both the activation timing of the Autonomous Emergency Steering (AES) maneuver and the geometric parameters of a quintic Bézier avoidance trajectory. Based on these parameters, a smooth and kinematically feasible reference path is generated for obstacle avoidance. For trajectory tracking, an MPC controller based on a dynamic vehicle model is implemented to follow the planned path while maintaining lateral stability. Simulation results in a static-obstacle scenario demonstrate that the proposed framework can generate feasible avoidance trajectories and track them with bounded steering response.
