Perception and Deep Reinforcement Learning-Based Local Path Planner for Autonomous Vehicles

Abstract

Autonomous driving has become a revolutionary area in modern transportation studies. End-to-end approaches offer an integrated paradigm that unifies perception, prediction, and planning functionalities within a single framework. This holistic approach streamlines the overall decision-making process, potentially improving the efficiency and accuracy of autonomous vehicle operations. In this paper, we propose an end-to-end reinforcement learning-based framework based on refined observations collected from sensors. Besides, the ray casting algorithm is introduced based on the top-view segmented images to generate the surrounding relative distances to the outset, which facilitates the surrounding distance parsing of an autonomous car. Subsequently, these distances, latent information from the driver’s view, and other sensor information are fed as an observation for a deep reinforcement framework to predict a steering angle. The experimental results showcase the effectiveness of methods implemented in self-driving vehicles within CARLA simulation environments.