Reinforcement Learning Paradigms in Autonomous Driving: A Perspective of Driving Tasks and Design Choices

Reinforcement Learning (RL) has emerged as a promising paradigm for autonomous driving (AD) thanks to its ability to learn policies through dynamic interactions with complex driving environments. Recent advancements in RL-based AD have progressed beyond standalone RL algorithms toward a diverse set of extended and hybrid learning paradigms, tailored to specific design choices across driving tasks. However, existing surveys have yet to categorize and analyze how these paradigms shape design decisions and performance trade-offs across the AD stack. In this survey, we address this gap by providing a systematic understanding of the most recent developments in autonomous driving, scrutinizing through the dual lenses of driving tasks and learning paradigms. We establish an explicit mapping between paradigm shifts, architectural components, and algorithmic solutions designed to overcome the shortcomings of conventional RL-based autonomous driving. Through an in-depth review of works predominantly published between 2023-2025, we analyze how major paradigms address the key limitations of classical deep RL, discuss open challenges as new hybrid paradigms arise, and motivate the RL-based AD outlook. Beyond synthesizing the emerging trend, this paper identifies recurring design patterns, paradigm-specific evaluation practices, and underexplored directions, offering a structured reference for researchers, practitioners, and new entrants to the field.