Delay- and Dropout-Aware GRU-SAC Measurement Covariance Adaptation for Robust UWB/INS Indoor Navigation

Ultra-wideband (UWB) ranging provides absolute indoor navigation constraints, whereas inertial navigation (INS) provides high-rate propagation but drifts. Non-line-of-sight (NLOS) bias, packet dropout, and delay make fixed UWB measurement covariance unreliable. We propose a delay- and dropout-aware framework that adapts anchor-wise covariance within an error-state extended Kalman filter (ESKF). A gated recurrent unit encodes pre-update innovations, baseline normalized innovation squared (NIS), geometry, quality, motion, delay, availability, and reliability history; soft actor–critic outputs bounded covariance scales. A geometry-aware belief preserves per-anchor reliability through missing packets, while delay inflation and NIS-regularized training discourage stale or overconfident updates. In a severe 50 m scale simulation with NLOS, 50% dropout, and 200 ms delay, position RMSE decreased from 0.2443 to 0.0668 m (72.7%) and P95 error from 0.4624 to 0.1132 m (75.5%) relative to a fixed-covariance ESKF; the empirical NIS-gate hit rate increased from 43.1% to 91.8%. In two selected snippets from one four-anchor experiment, RMSE decreased from 0.219 to 0.046 m (79.0%) relative to the fixed ESKF. The physical study provides a configuration-specific feasibility evaluation using one platform, one four-anchor layout, and two selected trajectory snippets.