DDPG-Based Path Planning Approach for Autonomous Driving

The present work develops a DDPG-based path planning algorithm that integrates the artificial potential field method into reinforcement learning to learn and generate an obstacle-free path quickly and autonomously. The vehicle kinematic model is adopted to describe the motion of autonomous vehicles, and the potential field function of obstacles, road boundaries as well as reference waypoints are considered to construct rewards of reinforcement learning, which enables the vehicle to realize the tradeoff between avoiding obstacles, preventing driving off the road and following the reference route. In contrast to the existent path planning algorithms, the proposed approach is able to learn autonomously in different driving environments, which is more suitable to autonomous vehicles. Moreover, simulations are provided to further demonstrate the effectiveness and adaptability of the proposed algorithm.