Experimental Testing for a Learning-based Powered-Descent Guidance Algorithm

Feature-based Learning for Optimal Abort Guidance

“…A deep neural network (DNN) in [ 21 ] has been applied to map the relationship between the initial conditions and the optimal solution for the powered descent problem. Our prior work in [ 22 ] demonstrate the ability of the supervised learning method to solve an optimal control problem real time along with the onboard implementation for the powered descent guidance problem. In addition, image-based deep reinforcement learning has been applied for autonomous lunar landing [ 23 ].…”

“…The purpose of the experimental tests is to replicate the dynamics of an abort spacecraft that is originally planned to land on the Moon and then verify the computational performance of the abort guidance algorithm in a constructed scenario. Our work in [ 22 ] built a customized quadcopter to validate the computational performance of a fuel-optimal powered descent guidance algorithm. Although scale-model rocket powered vehicles have been developed for testing EDL missions [ 28 ], [ 29 ] with the purpose of increasing the technology readiness level (TRL), the complexities and costs involved in the scale-model tests make them not viable until the late stages of a mission.…”

“…The components used for the quadcopter are listed in Table 4. The detailed description of critical components and the functionality of the customized drone can be found in [ 22 ]. The labelled quadcopter components are shown in Fig.…”