Confidence-aware learning for real-time optimal terminal guidance with impact time constraints

Constrained real-time optimal terminal guidance is essential for high-precision strike missions of missiles, but is limited by onboard computational constraints. Data-driven supervised learning has therefore emerged as an alternative to numerical optimization by enabling fast online inference through offline training. However, its deployment for guidance problems is still limited by two challenges. First, learning-based guidance relies on large-scale optimal datasets, while existing numerical optimal control solvers often lack sufficient efficiency and stability to generate scalable and well-distributed training data. Second, most learning-based guidance networks operate as black-box models whose outputs strongly depend on training data distribution, making it difficult to guarantee reliability. To address these challenges, this paper proposes a confidence-aware learning framework for optimal terminal guidance with impact time constraints. An improved backward integration data generation method guided by the state transition matrix is developed to achieve efficient and uniform coverage of the desired state space. An error distribution smoothing method is further introduced to construct a lightweight dataset for Gaussian process regression with uncertainty quantification. Finally, a confidence-weighted combined guidance strategy is designed by adaptively fusing the learned policy with an analytical guidance law. Numerical simulations demonstrate the optimality and reliability of the proposed approach for constrained terminal guidance.