Coupled path-attitude optimization for low-observable flight in complex terrain via differentiable neural surrogates

초록

This paper presents a coupled path-attitude optimization framework for low-observable UAV flight in complex terrain under radar and surface-to-air missile (SAM) threats. The method combines terrain masking, radar signalto-noise ratio (SNR) reduction, and dynamic weapon engagement zone (WEZ) avoidance in a single trajectory design problem. To represent the strong nonlinear dependence of detection risk on vehicle position, attitude, and terrain geometry, multilayer perceptron surrogates are constructed for the radar SNR and a terrain exposure index. The surrogate models provide differentiable approximations that can be embedded in a sequential convex programming (SCP) loop, while automatic differentiation is used to obtain the Jacobians required for linearization. To reduce the impact of false-negative visibility predictions near exposure boundaries, a conditional value-at-risk (CVaR)-based safety margin is introduced in the terrain exposure model. The resulting optimization problem includes nonlinear flight dynamics, surrogate-based detection metrics, terrain clearance, and dynamic WEZ constraints. Simulation studies in a mountainous environment show that the framework generates feasible trajectories that avoid the prescribed WEZs and exploit terrain occlusion when available. In exposed segments, the optimizer also selects vehicle attitudes associated with lower surrogate-predicted SNR, indicating that position and attitude can be coordinated to reduce the predicted radar detection metric within the proposed surrogate-based formulation. Weight-sweep results further illustrate the trade-off between terrain masking preference and SNR reduction.

키워드