Real-time guidance for powered landing of reusable rockets via deep learning

This paper focuses on improving the autonomy and efficiency of fuel-optimal powered landing guidance for reusable rockets considering aerodynamic forces. Deep-learning-based methods are developed to enable online and autonomous operation ability and to avoid the convergence problem encountered by classic indirect and direct optimal control methods. Considering the complex uncertainties of the preceding entry flight and potential unsettling hand-over conditions, a classification network is designed to classify the initial states of landing flights into different categories that correspond to bang-bang or non-bang-bang/singular thrust profiles. Thus, the subsequent online regression network can perform well for a large initial state distribution, and the algorithm adjusts to extensive landing situations. The combined application of classification and regression networks is one of the main contributions of the paper. The offline trained state-action regression networks generate guidance commands according to the real-time rocket state, obtaining a near-optimal landing trajectory. In addition, an online parallel trajectory simulation strategy is proposed to verify the trajectory quality, and an alternative trajectory optimization procedure is embedded into the proposed network-based framework to enhance the safety and accuracy of the guidance algorithm, representing another major contribution. Numerical experiments are presented to evaluate the effectiveness and accuracy of the proposed algorithm.