Neural convex optimization for real-time trajectory generation of asteroid landings

Abstract

This study presents a neural convex optimization framework to achieve high-efficiency and high-accuracy trajectory generation for asteroid landings. The core innovation lies in integrating a deep neural network (DNN) based gravity model into the recently developed reduced space sequential convex programming (rSCP), thereby leveraging the complementary strengths of both methodologies. Two specific neural rSCP methods are developed using fixed-point iteration and Newton-type iteration, both of which can convexify the DNN-related dynamics effectively, while avoiding the computational burden of DNN-related Jacobians. Specifically, the fixed-point iteration-based method approximates state-dependent terms with reference solutions from previous iterations, inherently eliminating the need for Jacobian computations. The Newton-type iteration-based method leverages inexact Jacobian information derived from inexact Newton iterations to circumvent the requirement for exact Jacobian computations. Numerical experiments on a fuel-optimal asteroid landing scenario validate the effectiveness and computational advantages of the proposed methods.