<p>Reinforcement Learning (RL) has significantly advanced the research topic of human-Artificial-Intelligence (AI) collaboration, offering powerful methods for improving the interaction between humans and AI systems across a range of applications. Despite the rapid growth of research on this topic, a review systematizing the challenges, mechanisms, and methods specifically tailored for human-AI collaboration is still lacking in the literature. This review focuses on RL-based human-AI collaboration, which can be structured into the following four parts: (1) comprehensively summarizing and analyzing the main challenges in human-AI collaboration, including aligning AI systems with human behaviors, generalizing to new human collaborators, ensuring robustness in dynamic environments, and performing coordination among multiple agents; (2) discussing the core mechanisms to address these challenges, including behavior characterization, intention understanding, and multi-agent coordination, and presenting a human-centric viewpoint on these mechanisms; (3) exploring the methods containing each mechanism, and investigating their effectiveness in enhancing human-AI collaboration; (4) comparing the representative methods from the perspectives of scalability, adaptability, and interpretability, and identifying the current challenges haunting this topic and proposing the potential directions for the prospective research on it. We expect this work will provide valuable insights into the academic advancements and inspire the technological breakthroughs in RL-based human-AI collaboration in the near future.</p>