ESA GNC Conference Papers Repository

This manuscript explores the concept of model-learning predictive control for orbit-attitude station-keeping in the vicinity of an asteroid. The asteroid gravity field inhomogeneities are assumed to be unknown a priori. In order to infer the gravity model parameters, these are simultaneously estimated with the state through an unscented Kalman filter. The progressive gravity model identification is combined with a model-learning predictive control strategy. The predictive control scheme increases its accuracy since the model parameters are estimated in-situ. Consequently, the tracking errors decrease over time as the model accuracy increases. Numerical results are shown and discussed, comparing the learning-based MPC strategy to a more classical non-learning approach to demonstrate the benefits and trade-offs of the former with respect to the latter.