AI Infers Cislunar Orbits from Angles-Only Measurements

This research significantly advances the field of generative astrodynamics by applying generative modeling to the complex problem of orbit determination in the cislunar environment. The specific challenge addressed is inferring the initial state's probability distribution using only angles-only measurements taken over short observation periods. The proposed solution involves training a conditional normalizing flow on perturbed topocentric observations, specifically from Near Rectilinear Halo Orbits. This approach allows for the creation of a flexible and potentially multimodal representation of the posterior distribution, which is crucial for accurately capturing the uncertainties inherent in such measurements. When new measurements are introduced, the learned density can be sampled to generate statistically consistent and physics-informed hypotheses about the spacecraft's state. These AI-generated estimates then serve as competitive "warm starts" for traditional nonlinear least-squares minimization algorithms, significantly improving their efficiency and accuracy in refining the orbital predictions.