GMM-Guided Neural MHE for Robust UWB-Based Robot Localization

Abstract

The ultrawideband (UWB) sensors are widely used for mobile robot localization in indoor and GNSS-denied environments. However, practical UWB measurements are often affected by nonline-of-sight propagation, multipath reflections, and intermittent outliers, which may introduce non-Gaussian residuals and degrade sensing reliability. To improve robustness under unreliable UWB measurements, this article proposes a Gaussian mixture model (GMM)-guided neural moving horizon estimation (MHE) framework for mobile robot localization. The GMM is used to model the residual distribution and estimate the posterior outlier probability of each UWB residual. A confidence factor is then constructed to attenuate unreliable measurement information in the MHE cost. Meanwhile, a lightweight neural network is introduced to adaptively adjust the weighting matrices in the MHE objective. In this way, probabilistic UWB reliability evaluation and adaptive MHE weighting are integrated within a unified estimation framework. The real-world LOS and NLOS experiments demonstrate that GMM-NNMHE improves localization accuracy and suppresses large positioning errors compared with filtering, fixed-weight MHE, neural MHE, and robust factor-graph baselines.