<p>Point cloud registration is a fundamental problem in computer vision, aiming to find the optimal transformation to align a pair of point clouds. In general, better point selection can improve the performance of registration, but there is still relatively little relevant research. To address this problem, a Fine Reinforcement Learning model with Trusted Point Selection (FRLTP) is proposed for point cloud registration in this paper. The model as a whole adopts a reinforcement learning framework that defines transformations including rotations and translations as the action space, joint embedding of source and target point clouds as the system state, and defines fused reward that are used to iteratively train the system along with PPO reward. The main innovation of the model is to utilize graph convolutional network (GCN), cross-attention in the state embedding module to initially obtain the information of the points and then utilize the similarity matrix to filter the key points. In order to test the effectiveness of the proposed model, experiments were conducted by applying it on benchmark datasets with partial overlap and noise. Numerical results show that our proposed model demonstrates consistently strong performance compared to the SOTA methods across multiple benchmarks. Ablation experiments show that the proposed Trusted Point Selection sub-module and fused reward are effective.</p>