Improving Observability of Relative Orbit Estimation Using Bearing Measurements and Light Curves

What the math gives to ML

The paper provides a concrete observability mechanism: bearing-only measurements leave relative range weakly identifiable, while photometric light-curve measurements add range- and geometry-dependent sensitivity, improving the Fisher information matrix and estimator convergence. This transfers naturally to latent-state neural networks, especially recurrent state-space models and world models, by training latent dynamics so that their predicted multimodal observations produce a well-conditioned Fisher information matrix. The most useful implementation is an observability regularizer that maximizes information over short rollout windows, with a quantitative prediction that adding the photometric auxiliary channel increases the smallest singular value of the latent observation Jacobian and reduces long-horizon state uncertainty.