Probabilistic Condition-Aware Dual-Branch Model for Multipath-Resilient Wireless Localization

Abstract

Wireless localization in complex propagation environments remains challenging due to the heterogeneous channel-location relationships induced by varying propagation conditions. Conventional unified localization models often fail to adequately capture such condition-dependent characteristics, leading to degraded positioning accuracy and limited robustness. To address this issue, this letter proposes a probabilistic condition-aware dual-branch localization framework that explicitly incorporates propagation uncertainty into the localization process. By modeling the location posterior as a mixture of LOS- and NLOS-conditioned predictors and adaptively fusing their outputs via a learned probabilistic router, the proposed approach enables propagation-aware localization without requiring explicit state observation at inference. A two-stage training strategy ensures stable learning of state-specialized representations and routing weights. Experiments demonstrate that the proposed method consistently outperforms both unified regression and hard-decision baselines, particularly in challenging NLOS scenarios. Source code is available at https://github.com/jzengust/Res-SR.