Validating Online Guidance for Reusability Flight Experiment: Genetic Programming and Reinforcement Learning

Kurzfassung

In this study, two machine learning (ML) techniques-genetic programming (GP) and deep reinforcement learning (DRL)-are leveraged to derive a reentry guidance law and are evaluated using a high-fidelity six-degree-of-freedom simulator developed for validating the Reusability Flight Experiment (ReFEx) vehicle guidance, navigation, and control (GNC) subsystem. Both methods are benchmarked against each other and against the baseline ReFEx optimization-based guidance strategy to assess their applicability to a real mission and to understand their respective strengths and weaknesses. The ReFEx mission focuses on the reentry phase, wherein GP and DRL models are applied to generate real-time corrections to precomputed reference guidance commands, thereby compensating for external disturbances and model inaccuracies. GP is selected for its ability to produce human-readable, continuous, and differentiable models that yield smooth guidance commands, whereas DRL employs a fully connected neural network (NN), delivering superior performance at the expense of a black-box model and nonsmooth guidance signals. The results demonstrate that DRL and GP achieve performance comparable to the mission baseline guidance, hence validating their applicability in mission-grade applications. Moreover, both ML approaches achieve faster online execution times than the baseline guidance method.