Enhancing Cislunar Operations by Integrating Reinforcement Learning into Classical Relative Guidance Methods

The growing interest in cislunar exploration demands guidance systems capable of handling the nonlinear dynamics of Near-Rectilinear Halo Orbits (NRHOs).Ensuring autonomy and minimizing human intervention are essential in such deep space missions, where communication delays impose strict self-sufficiency.This work investigates the integration of Reinforcement Learning (RL) into classical relative guidance strategies to enhance efficiency, flexibility and robustness during proximity operations without compromising safety and interpretability of the system.The proposed algorithm employs a hybrid approach: Approximating Sequence of Riccati Equations (ASRE) is used for trajectory optimization, Artificial Potential Field (APF) ensures path constraint handling, and Sliding Mode Control (SMC) provides robust control.The RL agent operates at a high level, selecting among predefined guidance configurations based on system observations.The framework is validated through extensive simulations in a CR3BP environment, focusing on rendezvous and docking in different scenarios.Results show that the RL-aided system outperforms traditional APF+SMC configurations, with significant improvements in propellant efficiency, Time Of Flight (TOF) and decision adaptability, while maintaining operational safety even under partial system failure.The proposed structure offers a scalable and modular solution to future autonomous guidance challenges in complex environments.