Kinematics-informed neural network control on SO(3)

Abstract

This paper presents an adaptive geometric control method for dynamic-model-free attitude tracking on the manifold of 3D rotations (SO(3)). Utilizing well-established definitions of attitude errors on SO(3), we develop a general control-affine linear error system. The input to this system is implicitly approximated by a kinematics-informed neural network (NN), which serves as the controller. The weights of this NN, designed to be inherently bounded, are adjusted online using a modified gradient descent strategy that relies solely on system kinematics. We demonstrate the effectiveness and online learning capability of our proposed method through comprehensive simulation results, using a satellite attitude control system as an example. A comparative analysis is also provided to validate our approach.