Multi-stage trajectory planning method for hazard zone avoidance under uncertainty based on neural networks

The trajectory planning method based on optimal control can maximize the flight capability of ultra-long-range glide-guided projectiles. However, when faced with uncertain battlefield conditions, this numerical method, which is time-consuming and prone to poor convergence, makes online applications difficult. To address this issue, a deep neural network-based trajectory planning method was proposed. This method utilized the non-linear mapping ability of deep neural networks to approximate the pseudospectral method computation model, reducing the computational load on the onboard computer of the projectile. The implementation of the method was primarily divided into two steps. The first step connected pre-contact points in three-dimensional space based on continuity conditions, considering various random states of the projectile and environment, using the multi-stage Gaussian pseudospectral method (MGPM), forming a range-optimal trajectory sample database that satisfies path constraints. The second step involved mapping out the optimal trajectory planning model by having the deep neural network learn the optimal actions of the projectile under different states offline, based on the optimal trajectory data sample library. Simulation results show that the proposed method can quickly generate near-optimal trajectories under random, demonstrating good real-time performance and robustness, making it suitable for solving online trajectory planning problems.