UAV State Estimation Using Physics-Informed Neural Networks

This paper presents a framework for state estimation of a fixed-wing Unmanned Aerial Vehicle (UAV) using Physics-Informed Neural Networks (PINNs). Unlike conventional data-driven or observer-based methods, PINNs combine partial measurement data with the UAV’s governing dynamics to form a unified loss function that simultaneously enforces measurement consistency and the underlying physical laws. The paper derived the 6-DOF (six degree-of freedom) rigid-body equations of motion for a fixed-wing platform and reformulated them into a general PDE/ODE residual. A feedforward neural network is then trained to match available sensor data while ensuring the neural network’s output satisfies the UAV’s kinematic and dynamic equations as a soft constraint. Numerical experiments highlight reduced data requirements, robustness to sensor noise, and improved fidelity in unobserved states compared to purely data-driven approaches.