Multi-Modal Sensing-Assisted Cross-Environment Beam Prediction Towards 6G

Abstract

Multi-modal sensing has demonstrated remarkable efficacy for intra-site beam prediction. However, generalizing trained models to unseen environments remains a significant challenge, primarily due to the inherent environment-specific wireless propagation characteristics. Existing cross-environment works rely on either ray-tracing simulations or single-modality data, leaving the real-world multi-modal setting largely unexplored. In this paper, we establish the first systematic cross-environment beam prediction benchmark on the real-world dataset and identify label-space shift as the underlying cause of poor cross-environment generalization, where different antenna orientations assign different physical meanings to the same beam index. Through comprehensive evaluation of 10 domain adaptation methods, we show that the adaptation methods tailored for feature-distribution shift are mismatched to this problem under the evaluated scenarios. Building on these insights, we propose a Modality-Aware Ranking-guided Adaptation (MARA) method, which addresses the task-specific challenges through physics-informed modality-aware learning rates and a pairwise beam ranking loss that exploits full power vector measurements. The experimental results demonstrate that the proposed MARA method significantly outperforms the few-shot fine-tuning baseline, improving distance-based accuracy (DBA) score by 6.6% at 100 adaptation samples.