Neural Observer-Based Learning for Robust State Estimation

This work introduces a state estimation approach that combines physics-based modeling with data-driven learning. The proposed estimator combines a PI observer and neural network to improve learning for the unknown parts of the system using gradient-based methods. The gradient is obtained using sensitivity propagation. Additionally, we develop a model-based policy gradient algorithm to learn directly from trajectory tracking errors without the need for ground truth. To enhance adaptiveness, a framework is developed to continually improve the accuracy of the dynamic model during deployment. Finally, we demonstrate the effectiveness of the proposed method through the simulation of vehicle tracking trajectories.