Uranus Orbiter Guidance: Comparing Numerical Predictive and Machine Learning Approaches for Aerocapture

Aerocapture is essential for enabling ambitious missions such as the proposed Uranus Orbiter and Probe, a flagship-class mission prioritized by the Planetary Science Decadal Survey. This paper presents a comparative analysis of data-driven guidance strategies for real-time aerocapture at Uranus, focusing on bank angle modulation. Specifically, we compared Feedforward Neural Networks, Generalized Additive Models, and Reinforcement Learning against the traditional fully numerical predictor-corrector aerocapture guidance methods. Results demonstrate that while traditional numerical methods maintain high accuracy under nominal and dispersed conditions, Machine Learning (ML) significantly reduces computational runtime, which is critical for onboard guidance algorithms. While the current ML implementation exhibits reduced accuracy in edge case scenarios, this work highlights the potential of data-driven approaches to address the demanding real-time control requirements of future space missions with hypersonic atmospheric guidance. This work presents mitigating these edge case inaccuracies through refined reward function design and improved training datasets.