<p>Ensuring accurate keypoint estimation in gesture point clouds is crucial, and to achieve this, it is imperative to fully leverage the detailed geometric information inherent within hand point clouds. This paper proposes a novel and efficient approach to 3D hand pose estimation based on a relational graph attention mechanism, designed to model the complex dependencies of hand gestures, which vary significantly in shape, orientation, and trajectory. By constructing a relational graph over the hand point cloud and dynamically adjusting attention weights according to node interactions, the model enables comprehensive feature extraction while enhancing interpretability through transparent visualization of keypoint dependencies. Furthermore, considering that 3D hand structures primarily consist of curves and surfaces, we introduce a local feature aggregation strategy to effectively capture fine-grained geometric features, leading to more precise representations of hand shape and structure. Extensive experiments on the MSRA, ICVL, and NYU datasets show that our method achieves average keypoint estimation errors of 7.3 mm, 6.1 mm, and 9.7 mm, respectively. Furthermore, the proposed model operates with only 14.3 MB of storage and achieves a real-time inference speed of 96.7 fps, significantly surpassing recent state-of-the-art methods such as HandDiff in both computational efficiency and deployability. Ablation studies further confirm the effectiveness of the relational graph attention mechanism and local feature aggregation strategy, validating the robustness, efficiency, and interpretability of our approach for 3D hand gesture recognition tasks.</p>