Offline-online hybrid attitude control for reentry vehicle: a continuous learning strategy via dual-timescale meta-adaptation

Reentry vehicles face dual challenges of persistent model uncertainties and time-varying actuator degradation–phenomena requiring fundamentally different compensation strategies yet occurring concurrently during critical mission phases. Conventional adaptive methods are insufficient because one-shot reentry missions lack offline fault data for training, whereas existing online learning techniques often overlook the intrinsic structure of rotational dynamics. This work establishes a continuous learning framework for aerospace attitude control where fault compensation progressively improves across repeated episodes through accumulated cross-episode knowledge. The framework integrates an offline-trained gradient-enhanced physics-informed neural network baseline with an online dual-timescale meta-learning compensator. The core innovation lies in moment-based architecture realizing continuous learning through temporal separation: the inner loop rapidly adapts task-specific parameters for immediate fault response, while the outer loop accumulates cross-task meta-knowledge through progressive encounters. The moment-based formulation learns compensation moments directly from inertial residuals via angular momentum conservation, substantially reducing parameter dimensionality while preserving complete physical information. Simulation validation demonstrates continuous learning through progressive improvement across episodes, robust performance under combined disturbances, and superiority over conventional methods, demonstrating that continuous learning offers a practical and robust solution for fault-tolerant aerospace.