Missile Autopilot Design via Neural Koopman Embedding

A learning-based missile autopilot design scheme is proposed, leveraging deep neural networks and Koopman operator theory to facilitate the missile control design process. The proposed method trains a neural Koopman embedding, a deep neural network that approximates the state of the embedded bilinear system for nonlinear missile dynamics using trajectory data. This Koopman-embedded bilinear model enables the systematic design of a state-dependent Riccati equation-based controller with local stability guarantees. The proposed method facilitates a full-envelope missile control design without the intricate synthesis of multiple controllers required for gain scheduling. Numerical simulations validate the effectiveness of the approach, demonstrating its tracking performance and robustness in handling nonlinear missile dynamics.