Abstract: Aiming at the insufficient accuracy and the low efficiency of 3D point cloud registration, a point cloud registration method based on subgraph matching and reinforcement learning was proposed to achieve the accurate and fast registration of low-quality point cloud. Firstly, the 3D point cloud registration can result from a series of discrete rigid transformation actions, and this work used a reinforcement learning strategy to train an end-to-end model to iteratively predict the rigid transformation actions. Then, for the model architecture, a Siamese backbone was used to extract the local feature information of the source point cloud and the target point cloud, respectively. Similar nodes in the source graph and the target graph were associated through a proposed cross-graph attention module. The aggregation of graph nodes was designed to extract the global features of two graphs, by using the weighted sum with gating vectors. Finally, the global features of the source graph and the target graph were fused, and the discrete rigid transformation action was predicted based on the fused feature. The reinforcement learning strategy significantly improves the generalization of point cloud registration. The cross-graph attention module further improves the accuracy and efficiency of point cloud registration. Extensive experiments on both synthetic and real-scanned datasets, ModelNet40 and ScanObjectNN, demonstrate that compared with the latest point cloud registration method, ReAgent, the proposed method can reduce the mean average error of rotation by at least 0.16 and the isotropic rotation error by at least 0.16, effectively improving the accuracy of registration on low-quality point clouds.