Comparison between Regressive and Classifying Neural Networks for PID Controlled Path-Following

In this paper, a regressive CNN-based system is developed with the objective of overcoming the turning angle limitations of a previously developed path-following classifying neural network. Images from a forward facing camera are put through a CNN and the result is supplied to a PID controller to guide the robot down a path. Trials were run for both current and previous methods and their results are compared and contrasted, with the new method completing all courses. Our regressive CNN was determined to have a RMSE of 3.584 and a R² of 0.987.

I. Introduction

Autonomous path-following is an important ability applied by autonomous guided vehicles (AGVs) and self-driving vehicles (SDVs). AGVs are employed extensively by the warehousing and manufacturing industries to lower operational costs and improve production efficiency [1], with modern guidance techniques for AGVs involving either fixed paths or computer-determined paths [2]. Unlike AGVs, fully autonomous SDVs are still in their research and development stage [3]. Once improvements in the developments of SDVs allow them to be commercially available, their anticipated advantages include the environmental benefits of reducing fuel consumption, air pollution, and carbon emissions, and the public safety advantage of reducing traffic accidents [4].