Line-of-Sight Guidance: Learning to Look Ahead

This article introduces a novel line-of-sight (LOS) guidance approach utilizing a neural network to parameterize the lookahead distance. We prove that the proposed guidance law renders a compact set containing the origin uniformly globally asymptotically stable (UGAS) for the closed-loop system. Importantly, if the sideways velocity is zero, the proposed guidance law renders the origin of the closed-loop system UGAS. By employing a neural network to parameterize the lookahead distance, the learning process results in a locally optimal lookahead distance for a given performance metric, and allows for nonlinear variation of the lookahead distance based on arbitrary input. We demonstrate how the proposed approach outperforms several state-of-the-art LOS guidance schemes utilizing time-varying lookahead distances.