Reinforcement learning-driven adaptive covariance control for robust automated INS/UWB navigation

This paper proposes a reinforcement learning (RL)-based adaptive covariance scaling framework for robust INS/UWB integrated navigation in dynamic and NLOS-prone environments. By formulating covariance tuning as a Partially Observable Markov Decision Process (POMDP) and employing a recurrent PPO policy, the method enables anchor-wise adjustment of the UWB measurement noise to balance accuracy and statistical consistency. Simulation results show that the proposed approach achieves an RMSE of 0.258m, outperforming classical adaptive filters and existing RL baselines. Real-world quadrotor experiments further demonstrate centimeter-level accuracy (0.036m RMSE) and strong robustness under severe NLOS and anchor dropout conditions, highlighting its effectiveness for resilient intelligent navigation systems.