Near-optimal 3D trajectory design in presence of obstacles: A convolutional neural network approach

Abstract

This paper proposes an approach based on neural networks for designing near-optimal 3D trajectories connecting two points separated by obstacles. A reference path is first built with a novel Theta* algorithm implementation, upgraded to reduce the number of waypoints and the angular variations. Starting from the path, a trajectory based on piecewise Bézier curves is designed with an algorithm relying on two design parameters. The optimal value of these parameters is estimated with a Convolutional Neural Network (CNN) architecture receiving as inputs an image of the path and its properties. The CNN training is performed on a synthetic dataset of optimal parameters built with Differential Evolution optimization for a variety of randomly generated paths. The parameters from CNN are refined with an algorithm capable of providing with high success rate near-optimal trajectories within minimum computation time.