Quasi-Optimal Learning based Analytical Powered Descent Guidance with State and Action Constraints

This paper proposes a Quasi-Optimal Learning based Powered Descent Guidance (QOLPDG) framework that combines the analytical structure of the optimal feedback guidance with data-driven learning to ensure quasi-optimal fuel performance under state and action constraints. The guidance acceleration is parameterized using a function form inspired by optimal feedback guidance law. Unlike existing methods that estimate the remaining time of flight only at the start of the guidance phase, a surrogate learning-based model is developed to estimate it continuously along the trajectory. This model is trained using simulation-based optimization to ensure that the generated trajectories satisfy system constraints and selects a fuel-optimal solution for each initial condition from a set of trajectories defined by the chosen function class. The method is termed quasi-optimal because it finds the best solution only within the space of trajectories parameterized by the chosen acceleration function. The advantage of computing remaining time-of-flight continuously, rather than once at the start of the guidance phase, is especially demonstrated under perturbed conditions. Additionally, the proposed method increases the controllable region compared to the classical guidance algorithm.