Investigação do desempenho do planejador de trajetórias Motion Planning Networks

Abstract

This study aims to investigate the application of a Machine Learning-based motion planner for solving problems in which a path between two points must be planned, avoiding any collisions with obstacles present in the environment. The environments herein used are bidimensional, static and feature obstacles of the same size. The chosen approach for solving the said problem was the Motion Planning Networks algorithm. In this planning algorithm, the following elements are highlighted: a neural network that synthesizes the environment data in an embedded and efficient representation; a neural network that gradually processes a set of candidate points to the planned path; an optimizer algorithm, which removes unnecessary points from such set; and, a replanning algorithm which attempts to fix eventual mistakes in the planned path. The viability of this approach was evaluated through the analysis of the following parameters: success rate in planning a valid path and computation time. As a result of said analysis, it is concluded that the Motion Planning Networks algorithm is consistent in planning valid paths for said problems, rarely failing; besides, it regularly took less than half a second to plan a path. MPNet’s mean computation time is significantly lower than the one obtained using the RRT* planner for the same paths.