An RNN-Based Adaptive Generalized Maximum Correntropy Kalman Filter for UAV Cooperative Localization

Unmanned aerial vehicles (UAVs) play a crucial role in various critical applications, where precise and reliable navigation is essential. However, dynamic maneuvers and complex environments often introduce non-Gaussian noise (NGN) into sensor measurements, which degrades the performance of traditional Kalman filters. To address this, this paper presents an adaptive generalized maximum correntropy Kalman filter (GM-CKF) enhanced by a recurrent neural network (RNN), referred to as RA-GMCKF, for robust UAV cooperative localization. Our approach employs an RNN-based predictor to learn the mapping between historical filtering error features and optimal kernel parameters. Specifically, time series of error feature matrices are constructed from filtering data and used to train the RNN by minimizing the position root mean square error (PRMSE). The trained RNN then predicts kernel parameters for the GMCKF, enabling adaptive filtering in challenging NGN conditions. We evaluate the RA-GMCKF in a cooperative localization scenario involving UAVs equipped with inertial measurement units, Global Navigation Satellite System receivers, and ranging sensors. Simulation results demonstrate that RA-GMCKF achieves lower PRMSE and outperforms state-of-the-art methods under NGN conditions. These results highlight the effectiveness of data-driven kernel adaptation for robust UAV localization in complex noise environments.