<p>Point cloud registration involves estimating a rigid transformation to align a source point cloud with a target, a task essential in applications such as autonomous driving, robotics, and medical imaging. Recent advances in deep learning have significantly improved registration performance by providing robust and efficient solutions that address key limitations of traditional methods, including sensitivity to noise, outliers, and initialization. Despite these advances, a well-structured taxonomy of deep learning-based point cloud registration (DL-PCR) methods remains underdeveloped, limiting systematic comparison and analysis. This paper presents a comprehensive survey and taxonomy of DL-PCR. We begin with a formal definition of the registration problem and a review of commonly used datasets, evaluation metrics, and loss functions. Existing DL-PCR methods are then categorized by registration scenario–pairwise or multi-scan–and further classified by supervision strategy and methodological design. Supervised approaches are discussed in terms of registration procedures, optimization strategies, learning paradigms, and integration with traditional techniques, while unsupervised approaches are grouped into correspondence-based and correspondence-free methods. Multi-scan techniques are divided into multiview registration, which focuses on global pose estimation across multiple scans, and multi-instance registration, which involves jointly localizing and aligning multiple object instances. To support fair comparison, we provide quantitative evaluations of recent state-of-the-art methods under a unified training setup. Finally, we discuss open challenges and highlight potential research directions to guide future advancements in DL-PCR. A comprehensive collection of DL-PCR resources is available at <a href="https://github.com/yxzhang15/PCR">https://github.com/yxzhang15/PCR</a>.</p>