An Environmental Incremental Learning Framework for Vision-Aided Beam Prediction in 6G Networks

Vision based beam selection offers a compelling alternative to pilot based beamforming in millimeter wave (mmWave) communications, particularly within the dynamic operational contexts expected in sixth generation (6 G) wireless systems. By utilizing vision based sensing, such systems can perform rapid beam alignment while reducing dependence on frequent pilot transmissions. Nonetheless, current models are generally trained under fixed environmental settings and static data distributions which makes them unsuitable for real world deployments where conditions such as weather, lighting, and visual scene change over time. In this work, we propose an Environmental Incremental Learning (EIL) framework tailored for vision-guided beam prediction, which is capable of incrementally adapting to new environmental conditions such as clear daylight, rainfall, and snow-covered environments without requiring complete retraining of the model. Our method combines class balanced sample replay with partial knowledge distillation using previously trained teacher networks. Environmental changes are treated as sequential tasks, and the learning process jointly minimizes a supervised objective on new samples and a divergence based alignment on replayed instances. This formulation mitigates catastrophic forgetting while supporting scalable adaptation. Experimental evaluations reveal that the proposed method achieves higher Top 1 prediction accuracies, outperforming both replay buffer and fine tuning baselines. Additionally, it lowers average power loss by $61.33 \%$ to $88.57 \%$ compared to the different baselines. These outcomes highlight the proposed framework’s effectiveness in delivering resilient beamforming under evolving environmental conditions in future 6 G systems.