Neural Network-based Imitation Learning for Optimal Satellite Formation Control

Abstract

Satellite Formation Control (SFC) with on-off impulsive thrusters represents a significant challenge in space missions. Classic and optimal control approaches with control action quantization provide possible strategies, but limited in performance due to system linearity assumptions and model accuracy requirements. This paper investigates the use of a neural network controller synthesized via imitation learning to replicate the behavior of an expert Linear Quadratic Regulator (LQR) controller with quantized actions. A Multi-Layer Perceptron (MLP) neural network is trained on data generated by a LQR controller designed for 6U CubeSats formation with on-off thrusters, learning the state-action mapping that implicitly includes quantization effects. The goal is to assess whether a neural controller can achieve comparable performance to the expert while offering potential advantages in generalization and adaptability.