Adaptation of Noise Covariance in Extended Kalman Filter Using Reinforcement Learning for Improved UAV Attitude Estimation

Accurate attitude determination of unmanned aerial vehicles (UAVs) is crucial for autonomous navigation, particularly when relying solely on gyroscope, accelerometer, and magnetometer measurements without utilizing the Global Positioning System (GPS). Reinforcement learning (RL) has emerged as a promising artificial intelligence technique applicable across various domains. This research introduces a novel approach that leverages RL to enhance the performance of the extended Kalman filter (EKF) in attitude estimation. The proposed method depends of RL which uses the Q-learning model and policy to find best solution to adjust autonomously the measurement noise covariance matrix within the EKF. By establishing a reward mechanism that incentivizes actions minimizing the prediction error relative to true measurements, the RL dynamically optimizes the measurement noise covariance matrix. This innovative integration of RL and EKF, referred to as RL-EKF, has been implemented and tested. Results demonstrate that RL-EKF significantly outperforms the traditional EKF, yielding marked improvements in attitude estimation accuracy. The improvement ratios showed that selected method is very effective in the field of attitude estimation.