Research on Path Planning Method Based on Deep Reinforcement Learning for Autonomous Driving

Abstract

This article introduces a mixed deep reinforcement learning path planning framework (DRL-PP) for self-driving cars. It's designed to be efficient, safe, and verifiable in complex, dynamic settings. The framework is built on layered action representation: the policy network gives out multi-step reference waypoints, and the lower level uses MPC/PID to make sure things are physically possible. The reward function is designed in a combined way and adds a distance-based cubic booster to amplify learning signals when getting close to the goal. The training strategy uses a dual-network setup, experience replay, and a cautious ε-cooling schedule to improve convergence stability and sample usage. During operation, it adds Action Shield and constraint projections based on control barrier functions or quadratic programming for real-time safety fixes. Compared with baseline systems like DQN and Actor–Critic on a unified grid obstacle simulation platform, the proposed method shows clear advantages in success rate, crash rate, path smoothness, and sample efficiency (example results: 94.2% success rate, 3.1% crash rate, around 3.2×10^4 exploration iterations needed to reach stable performance).