Overview of Visual Reinforcement Learning Methods

Vision, as the primary means for reinforcement learning agents to perceive their environment, provides rich and detailed information that supports agents in making more complex and precise decisions. However, the high-dimensional nature of visual data often leads to information redundancy and low sample efficiency, posing a key challenge in the application of reinforcement learning. How to efficiently extract key visual representations from limited interaction data to enhance agents＇ decision-making capabilities has become a current research focus. To address this, this paper systematically reviews visual reinforcement learning methods, categorizing them into five categories based on their core ideas and implementation mechanisms: Image-enhanced, model-enhanced, task-assisted, knowledge-transferred, and offline visual reinforcement learning approaches. It provides an in-depth analysis of the research progress in each category, as well as the strengths and limitations of representative works. Meanwhile, this paper reviews four major benchmark platforms: DMControl, DMControl-GB, DCS, and RL-ViGen, and summarizes the applications of visual reinforcement learning in typical scenarios such as robotic control, autonomous driving, and multimodal large models. Finally, based on current research bottlenecks, this paper discusses future development trends and potential research directions, aiming to offer a clear technical framework and research reference for this field.