A tightly-coupled adaptive unscented Kalman filter for indoor mobile robot localization with UWB/odometry/gyroscope

In indoor environments, ultra-wideband (UWB) signals are susceptible to gross ranging errors caused by non-line-of-sight (NLOS) occlusion and dynamic robot motion. These errors significantly degrade the accuracy and robustness of conventional multi-sensor fusion localization methods. This paper proposes a tightly coupled UWB/odometer/single-axis gyroscope localization method based on an adaptive unscented Kalman filter (UKF) for industrial environments. The method constructs the robot kinematic equations and the tightly coupled UWB ranging observation equations within the nonlinear UKF framework, and incorporates an adaptive robust mechanism based on sliding-window residual statistics. The main contribution is a joint line-of-sight (LOS)/NLOS discrimination and adaptive gross error handling method that combines the instantaneous normalized innovation squared (NIS) with long-term historical residual consistency. When an observation is identified as NLOS, a soft rejection mechanism based on exponential confidence weights is adopted to suppress gross errors while retaining valid ranging information. Meanwhile, a long-term historical residual window is maintained in real time. The observation noise covariance is estimated online for LOS anchors and inflated for NLOS anchors, thereby adaptively suppressing the influence of unreliable observations during the filter update. Simulation experiments and real industrial tests are carried out to validate the proposed method, yielding average localization errors of 0.2046 m and 0.1699 m, corresponding to 11.12% and 2.74% reductions relative to the suboptimal Huber-UKF method, respectively. In real industrial scenarios with frequent dynamic maneuvers, the mean absolute error and root mean square error of the proposed method are 0.2282 m and 0.2559 m, which are reduced by 9.01% and 16.24%, respectively, while the average computational time increases by only 1.09 ms.