RS-CNN, or Relation-Shape Convolutional Neural Network, is proposed for point cloud analysis. It extends regular grid CNNs to irregular configurations, enabling the learning of high-level geometric relations among points. The key innovation is the relation-shape convolution, which learns a high-level relation expression from predefined geometric priors, allowing for explicit reasoning about the spatial layout of points. This leads to discriminative shape awareness and robustness. RS-CNN is designed as a hierarchical architecture that achieves contextual shape-aware learning for point cloud analysis. Extensive experiments on three tasks demonstrate that RS-CNN achieves state-of-the-art performance. The method is robust to point permutation and rigid transformations, and it effectively captures shape features from point clouds. The proposed approach outperforms existing methods in shape classification, part segmentation, and normal estimation tasks. The RS-CNN architecture is validated through comprehensive experiments and ablation studies, showing its effectiveness in learning shape-aware representations from point cloud data.RS-CNN, or Relation-Shape Convolutional Neural Network, is proposed for point cloud analysis. It extends regular grid CNNs to irregular configurations, enabling the learning of high-level geometric relations among points. The key innovation is the relation-shape convolution, which learns a high-level relation expression from predefined geometric priors, allowing for explicit reasoning about the spatial layout of points. This leads to discriminative shape awareness and robustness. RS-CNN is designed as a hierarchical architecture that achieves contextual shape-aware learning for point cloud analysis. Extensive experiments on three tasks demonstrate that RS-CNN achieves state-of-the-art performance. The method is robust to point permutation and rigid transformations, and it effectively captures shape features from point clouds. The proposed approach outperforms existing methods in shape classification, part segmentation, and normal estimation tasks. The RS-CNN architecture is validated through comprehensive experiments and ablation studies, showing its effectiveness in learning shape-aware representations from point cloud data.