Sliding Window Factor Graph-Based UWB/IMU/Wheel Encoder Fusion for Indoor Mobile Robot Localization

Accurate localization in GPS-denied indoor spaces remains difficult due to challenges in fusing ultra-wideband (UWB), inertial measurement unit (IMU), and wheel-encoder data, specifically anchor calibration and Non-Line-of-Sight (NLoS) mitigation. We present a unified framework to address these issues. First, a Geometric Dilution of Precision (GDOP)–aware global initialization of UWB anchors via semidefinite programming (SDP) achieves a 0.15m root-mean-square error (RMSE) in calibration. Next, a dual-stage NLoS mitigation—combining XGBoost-based link classification with Extended Kalman Filter (EKF) range correction—reduces the range RMSE from 0.43m to 0.12m. The core of our system is an Adaptive Huber Sliding- Window Factor-Graph Optimization (AHSW–FGO) method that tunes robustness online using recent residual statistics. In challenging indoor environments (corridor and office), AHSW–FGO achieves trajectory RMSE of 0.08m and 0.10m, respectively, improving upon standard EKF-based fusion by 50.0% and conventional sliding-window FGO by up to 38.5%. While reducing high-percentile errors, the method maintains a practical online latency of approximately 0.12 s per update. These results demonstrate that adaptive residual reweighting delivers substantial and consistent gains in both accuracy and robustness for indoor localization.