Abstract
Reentry trajectory optimization plays a vital role in the development of reusable spacecraft. However, the reentry process exhibits strong nonlinearities, characterized by parameter uncertainties and strict constraints. Consequently, the reentry trajectory optimization problem can be framed as an optimal control problem for a complex nonlinear dynamical system. Traditional trajectory optimization methods are often severely limited by slow convergence rates, limited robustness, and poor real-time performance, hindering the precise and efficient generation of control commands. To overcome these critical challenges, this study proposes a novel fast optimal robust trajectory generation method integrating polynomial chaos expansion, convex optimization, and deep neural network technologies. A large offline dataset of robust reentry trajectory optimization samples is constructed using nonintrusive polynomial chaos expansion and the sequential second-order cone programming method. Subsequently, a deep neural network is trained on this dataset to generate robust optimal control commands in real time. Compared to offline convex optimization methods, the proposed approach reduces the terminal state relative error to less than 2.1% and generates each bank angle command in 60 ms, reducing the total computing time by 97.8%. In conclusion, this novel trajectory generation method enables real-time, high-precision, and robust trajectory generation for reentry vehicles.