Reinforcement-Learning-Gated RTS Smoothing for Robust UWB-Based Train Trajectory Recovery Under Missing Measurements

Ultrawideband (UWB) ranging is a promising Internet of Things (IoT) sensing primitive for rail transit, because it can support infrastructure-assisted localization in global navigation satellite system (GNSS)-denied stations, tunnels, and platform areas. The main challenge considered in this work is not nominal positioning accuracy, but robust trajectory recovery under structured missing measurements caused by burst outages, correlated anchor blockage, and station-specific non-line-of-sight (NLOS) bias. We propose a reinforcement-learning-gated Rauch–Tung–Striebel (RTS) smoothing framework for UWB-based train trajectory recovery. Instead of applying the RTS backward correction unconditionally, we learn a bounded continuous gate that modulates the mean correction strength using observation-availability, innovation-consistency, and uncertainty-growth cues. The gate is optimized with soft actor–critic (SAC), while the Kalman/RTS backbone is preserved for interpretability and deployment controllability. The framework explicitly specifies the policy state, training protocol, and failure modes, and it adopts a mean-gated formulation to avoid the covariance-scaling ambiguity of naively gated RTS smoothing. Experiments on rail-style stopping trajectories with random, burst, geometry-related, and bias-induced degradations show that the proposed method improves station-region accuracy and tail-risk metrics over representative interpolation, machine-learning, and model-based baselines while introducing only a small inference overhead. These properties make the method attractive for practical IoT localization pipelines for stop-alignment scenarios.