This article reviews the current progress in semantic communications (SC), with a focus on the application of reinforcement learning (RL) within this field. SC enhances traditional communication by transmitting semantic information rather than complete data, thereby reducing bandwidth requirements while preserving the accuracy of the conveyed meaning. RL, a branch of machine learning, enables intelligent agents to learn from their actions and rewards in complex, dynamic environments. This paper not only reviews the theoretical foundations of SC and RL but also provides a comparative analysis of various RL approaches applied to SC, offering quantitative assessments of their performance in areas such as semantic similarity and transmission efficiency. We categorize and analyze existing research based on three primary dimensions of the SC system: the transmitter, which focuses on semantic extraction, encoding, and resource allocation; the channel, which ensures secure and efficient transmission of semantic information; and the receiver, which is responsible for semantic decoding, restoration, and multi-agent collaboration. Furthermore, we offer a balanced discussion on the advantages and potential improvements of various RL methods, providing insights into their suitability for different SC scenarios. Additionally, we discuss specific training strategies for RL agents in SC, covering exploration-exploitation trade-offs, data requirements, and adaptive learning approaches. Finally, we identify open issues in SC across various applications and scenarios, proposing potential directions for future research. By addressing these gaps, we aim to enhance understanding and simulate greater interest in further research in this emerging area.