Learning robust, distinctive, and generalizable 3D local features is crucial for point cloud registrationPoint cloud registration and other geometric processing tasks. Conventional methods often depend on handcrafted descriptorsHandcrafted descriptors susceptible to noise or deep learning models that fail to preserve rotation invarianceRotation invariance, leading to unreliable surface matchingSurface matching. This chapter presents SpinNet, a lightweight yet powerful neural networkNeural network designed to learn rotation-invariant local surface descriptors with high discriminabilityDiscriminability and generalizationGeneralization. The framework introduces a Spatial Point Transformer that maps the input surface onto a cylindrical representationCylindrical representation, ensuring SO(2) rotation equivariance and enabling end-to-end learning. To extract rich geometric structures, a Neural Feature Extractor is employed, combining point-based operations with 3D cylindrical convolutionCylindrical convolutions.