Neural network-based trajectory optimization for unmanned aerial vehicles

A direct trajectory optimization method that uses neural network approximation methods is presented. Neural networks are trained to approximate objective functions and vehicle dynamics. The neural network method reduces computational requirements by removing the need for collocation and providing fast computation of gradients when compared with direct and pseudospectral collocation methods. The method is shown to significantly reduce computational costs while resulting in trajectories comparable to those produced by direct collocation and pseudospectral methods. Because a neural network readily provides accurate computation of gradients, it removes the need for formulating analytical gradients; thus, the method is more easily extended to different types of applications with different objective functions and constraints. This paper demonstrates the flexibility of the neural network trajectory optimization approach through simulation of three cases: a single unmanned aerial vehicle operating a fixed camera, multiple unmanned aerial vehicles operating fixed cameras, and a single unmanned aerial vehicle operating a gimballed camera. The method's ability to produce tightly constrained trajectories is also demonstrated.