Robust radar object classification is a challenging task, primarily due to the aspect sensitivity limitation of one-dimensional High-Resolution Range Profile (HRRP) data. To address this, we propose Point-HRRP-Net. This multi-modal framework integrates HRRP with 3D LiDAR point clouds via a Bi-Directional Cross-Attention (Bi-CA) mechanism to enable deep feature interaction. Since paired real-world data is scarce, we constructed a high-fidelity simulation dataset to validate our approach. Experiments conducted under strict angular separation demonstrated that Point-HRRP-Net consistently outperformed single-modality baselines. Our results also verified the effectiveness of Dynamic Graph CNN (DGCNN) for feature extraction and highlighted the high inference speed and the potential of Mamba- based architectures for future efficient designs. Finally, this work validates the feasibility of the proposed approach in simulated environments, establishing a foundation for robust object classification in real-world scenarios.

High-Resolution Range Profile (HRRP)-based space object classification is severely limited by aspect sensitivity. Inspired by the intrinsic complementarity between HRRP and LiDAR point clouds, this work investigates the feasibility and effectiveness of fusing these two modalities to address this limitation. We propose the Point-HRRP-Net framework. This framework employs dual-stream extractors to independently encode HRRP electromagnetic signatures and 3D point cloud geometric topologies. Subsequently, a Bi-Directional Cross-Attention (Bi-CA) mechanism is designed to fuse the two modalities. To enable information interaction, this mechanism utilizes point-to-point attention to correlate radar scattering features with 3D geometric points, thereby constructing a comprehensive target representation. Due to data scarcity, we constructed a paired simulation dataset for evaluation. Experimental results demonstrate that the proposed framework consistently outperforms its constituent single-modality baselines. The model achieves 57.67% accuracy on the 180° split and demonstrates generalization capability to unseen viewpoints. Ablation studies further validate the efficacy of the Bi-CA mechanism and the selected feature extractors. Finally, we assess the potential sim-to-real discrepancies and evaluate deployment feasibility across various hardware platforms.