<p>Hand gesture recognition based on mmWave radar point clouds has excellent potential in the human-computer interaction field. However, conventional approaches often underutilise raw data and fail to incorporate the underlying physical properties of radar echoes, resulting in a poor trade-off between recognition accuracy and robustness. To overcome this limitation, we propose a Physics-Informed PointNet + + Network (PIPN). First, we integrate point cloud density and echo intensity to generate physics prior weights. Second, inter-frame dynamic correlations are calculated from the PointNet + + backbone’s feature extraction. Finally, we fuse the weights for density, intensity, and inter-frame correlation, embed the physical information of radar point clouds into the network’s feature learning, and guide the network’s attention to the most significant spatio-temporal regions. Evaluated on a comprehensive, self-collected mmWave radar dataset comprising 12 distinct gesture categories, our method enhances feature discriminability. It achieves state-of-the-art recognition accuracy of 99.89%, demonstrating a significant improvement over existing methods. Experimental results confirm the feasibility of our method. It proves a practical method for human-computer interaction.</p>