Onboard Optical Position Estimation for Autonomous Cislunar Navigation via Reinforcement Learning

Accurate onboard navigation is fundamental to spacecraft autonomy, especially in deep-space and cislunar environments where ground-based orbit estimation may introduce unacceptable latency. Optical navigation (OpNav) offers a viable solution, but conventional geometric methods typically require high-resolution imagery and substantial computational resources, limiting their applicability under challenging visual conditions and onboard hardware constraints. This paper presents a reinforcement learning (RL) framework for autonomous optical navigation in cislunar space. A convolutional neural network is trained to correct the position estimates by processing differences between simulated and observed lunar images. Training is performed in a simulated visual environment, enabling the policy to learn robust estimation strategies under observation noise, unmodeled dynamics, and varied initial conditions. The method is demonstrated for station-keeping along a southern halo orbit around the Earth-moon L2 point. Results show that the RL-based navigation policy consistently provides position estimates within the required accuracy for the closed-loop onboard controller to successfully maintain the spacecraft along the reference orbit, despite low-quality image inputs and a low update frequency. These findings underline the feasibility of RL-driven OpNav as a computationally efficient and resilient alternative to traditional techniques, offering a promising foundation for future onboard learning-based navigation systems in the vicinity of planetary bodies.