Robust Beam Prediction for V2X Networks with Multi-Modal Sensing

Summary

This paper proposes BeamTransFuser, a hierarchical Transformer architecture for beam prediction in V2X networks that progressively fuses camera, LiDAR, radar, and GPS observations collected at a road-side unit. Dedicated branches (ResNet34 for camera, ResNet16 for LiDAR and radar, an MLP for GPS) feed four-stage cross-modal attention blocks, and a softmax-weighted module combines modalities before an MLP selects one beam from a 64-beam codebook. For missing sensors, a CVAE-based generative module — trained offline with a fixed isotropic latent prior and fine-tuned with a beam-prediction loss — reconstructs absent modality features at inference. On DeepSense 6G urban scenarios 31–34 (11,243 samples, 90/10 split), BeamTransFuser attains an overall DBA-score of 0.9129 versus 0.8969 for ICMFE and 0.8910 for CMDF, with a perfect 1.0000 in Scenario 31, though QTNs (0.9124) and ICMFE (0.9074) stay ahead in Scenario 34, a gap the authors acknowledge. With simulated missing modalities, generation lifts Top-1 accuracy from 6.67% to 45.51% (missing radar) and 56.18% (missing LiDAR); the missing-camera case is omitted as too complex. Limitations: some baseline numbers are cited from prior papers rather than re-run under identical settings; no ablation isolates the hierarchical fusion design; results are single-run without variance or significance testing; and the latency advantage over diffusion models is asserted but not measured.