Benchmarking Transformer and Sequence Models for UWB Indoor Localization

Accurate indoor localization is a key enabler for many intelligent systems, including autonomous vehicles, service robots, smart infrastructure, and healthcare technologies. Ultra-Wideband (UWB) technology provides high-resolution range measurements and resilience to multipath, making it well-suited for such applications. However, in cluttered indoor environments, reflected signals under non-line-of-sight (NLoS) conditions can distort range estimates. These distortions are encoded in the Channel Impulse Response (CIR), which describes how signals propagate through the environment. This paper presents a comparative benchmark of learning-based models for predicting and correcting UWB ranging errors directly from CIR data. We evaluate three neural architectures—Transformer, BiLSTM, and CNN—within a unified pipeline and integrate their predictions into a classical Weighted Least Squares (WLS) localization algorithm via adaptive error weighting. Experiments on a publicly available UWB dataset spanning four diverse environments show that the Transformer-WLS model achieves the lowest average error in three out of four settings—including 1.08 m in an industrial space—outperforming BiLSTM-WLS (1.22 m), CNN-WLS (4.01 m), and traditional WLS ($\text{1. 3 3 m}$). These results highlight the effectiveness of attention-based sequence modeling for UWB localization and provide practical guidance for deploying robust learning-enhanced positioning systems in real-world environments.