Accurate angle-of-arrival (AoA) estimation with passive Ultra High Frequency (UHF) Radio-Frequency Identification (RFID) is a challenging task in realistic indoor settings due to multipath, phase ambiguity, and hardware-induced bias. In this work, we propose a physics-informed Bayesian sensor-fusion framework that treats AoA as a latent variable in a regression model. Classical array-processing estimates are incorporated as noisy physics-based observations and fused with feature-based regression, achieving calibrated uncertainty. The approach is validated with a fully commercial off-the-shelf (COTS) RFID system in an office environment. The best configuration obtains sub-degree accuracy, with root mean square error (RMSE) of 0.4993 ∘ and a mean absolute error (MAE) of 0.3699 ∘ , while also providing posterior predictive intervals that quantify confidence at inference time. These results demonstrate that combining lightweight physics models with Bayesian learning delivers reliable, accessible, and reproducible AoA estimation in multipath-rich environments using low-cost hardware.
Benelmekki et al. (Sun,) studied this question.