Integrated DNN-SMC Control With CLF-QP Stability Monitoring for Lunar Landing

Abstract

Optimal trajectory control problems for spacecraft landing often require real-time implementation, whereas traditional numerical optimization methods are difficult to deploy in time-critical onboard scenarios. Deep neural networks (DNN), owing to their universal approximation capability, have demonstrated strong potential in learning fuel-optimal landing trajectories from offline optimization data. Moreover, recent advances in specialized onboard computing hardware make real-time execution of trained neural networks increasingly feasible. Nevertheless, providing theoretical guarantees of closed-loop stability and robustness for DNN based guidance and control remains a major challenge. This paper proposes the integration of a DNN controller and a Sliding Mode Control (SMC) law into the objective function of the CLF-QP based stability monitor framework. A time-varying control policy is introduced to exploit the complementary strengths of both controllers: the DNN controller is primarily used in the initial phase where rich training data are available to capture optimality; in the intermediate phase, the DNN and SMC are blended to transition smoothly toward the terminal conditions; and in the final phase, where data scarcity and uncertainty dominate, the SMC alone is activated to guarantee robustness and stability. Simulation results for spacecraft landing demonstrate that the proposed method achieves near-optimal performance while maintaining provable stability throughout the descent.