Abstract

When intelligent spacecraft or space robots perform tasks in a complexenvironment, the controllable variables are usually not directly available andhave to be inferred from high-dimensional observable variables, such as outputsof neural networks or images. While the dynamics of these observations arehighly complex, the mechanisms behind them may be simple, which makes itpossible to regard them as latent dynamic systems. For control of latentdynamic systems, methods based on reinforcement learning suffer from sampleinefficiency and generalization problems. In this work, we propose anasymptotic tracking controller for latent dynamic systems. The latent variablesare related to the high-dimensional observations through an unknown nonlinearfunction. The dynamics are unknown but assumed to be affine nonlinear. Torealize asymptotic tracking, an identifiable latent dynamic model is learned torecover the latents and estimate the dynamics. This training process does notdepend on the goals or reference trajectories. Based on the learned model, weuse a manually designed feedback linearization controller to ensure theasymptotic tracking property of the closed-loop system. After considering fullycontrollable systems, the results are extended to the case that uncontrollableenvironmental latents exist. As an application, simulation experiments on alatent spacecraft attitude dynamic model are conducted to verify the proposedmethods, and the observation noise and control deviation are taken intoconsideration.