Abstract
Interstellar objects (ISOs) are likely representatives of primitive materialsinvaluable in understanding exoplanetary star systems. Due to their poorlyconstrained orbits with generally high inclinations and relative velocities,however, exploring ISOs with conventional human-in-the-loop approaches issignificantly challenging. This paper presents Neural-Rendezvous -- a deeplearning-based guidance and control framework for encountering fast-movingobjects, including ISOs, robustly, accurately, and autonomously in real time.It uses pointwise minimum norm tracking control on top of a guidance policymodeled by a spectrally-normalized deep neural network, where itshyperparameters are tuned with a loss function directly penalizing the MPCstate trajectory tracking error. We show that Neural-Rendezvous provides a highprobability exponential bound on the expected spacecraft delivery error, theproof of which leverages stochastic incremental stability analysis. Inparticular, it is used to construct a non-negative function with asupermartingale property, explicitly accounting for the ISO state uncertaintyand the local nature of nonlinear state estimation guarantees. In numericalsimulations, Neural-Rendezvous is demonstrated to satisfy the expected errorbound for 100 ISO candidates. This performance is also empirically validatedusing our spacecraft simulator and in high-conflict and distributed UAV swarmreconfiguration with up to 20 UAVs.