Real-time trajectory optimization for collision-free asteroid landing based on deep neural networks

Abstract

For the problem of autonomous and safe landing on irregularly shaped asteroids, a real-time collision-free trajectory optimization method based on deep neural networks (DNNs) is proposed. First, in order to overcome the difficulty of solving the optimal trajectories by indirect methods caused by anti-collision path constraints and the difficulty of establishing the mapping relationship between states and initial costates by DNNs, a two-stage anti-collision trajectory optimization model is constructed, which can provide the initial costates for trajectory optimization under anti-collision constraints of each stage through the approximate analytical solution of costates. It can efficiently generate a large number of optimal trajectory samples. Second, a region division strategy is proposed for further reducing the time-consuming of DNN training. Then, the path constraints are slightly relaxed to avoid the optimal trajectories predicted by DNNs from violating the anti-collision constraints. The proposed method is applied to the scenarios of landing on asteroids 101955 Bennu and 433 Eros. The simulation results indicate that this method can estimate the initial costates efficiently and accurately, and can plan the optimal trajectories of anti-collision landing in real time.