Optimal online trajectory generation for a flying robot for terrain following purposes using neural network

This paper is concerned with the problem of the two-dimensional optimal online trajectory generation for a flying robot for terrain following/terrain avoidance purposes using neural network. To this end, the terrain is modeled as a terrain following constraint. In the next step, the inverse dynamics method is utilized for offline trajectory generation. Since the main objective of this research is online trajectory generation, the neural network method is employed for this purpose. To train the neural network, various approaches such as steepest descent, conjugate gradient, resilient back propagation, and Levenberg-Marquardt are explored in detail, and finally, the Levenberg-Marquardt method is selected on account of some merits. The efficacy of the neural network method is demonstrated by extensive simulations, and in particular, it is verified that this method is able to produce a solution satisfying all hard constraints of the underlying problem.