Investigating the Impact of Variable Effects in Virtual Training on the Behavior of a Physical Autonomous Robot

Abstract

AbstractArtificial Intelligence is a current megatrend in computer science and almost every aspect of digitization. A scenario for training robots in a virtual environment to fulfil tasks in the real world is created to offer engineering students relevant insights into the field of neural networks. The proposed system generates training data to train a convolutional neural network (CNN) to autonomously drive a mobile robot using computer vision. The physical aspect of the setup includes the mobile robot and two different race circuits to evaluate the driving characteristics. The digital aspect consists of a 3D environment and a digital representation of the physical robot, both of which are developed using the game engine Unity. Inside the 3D environment, an infinite, procedurally generated road is created for the digital robot to drive on. The generated images from the virtual camera of the virtual robots are the basis to train the CNN to maneuver the physical robot in the real-world experiment.KeywordsMachine learningConvolutional neural networkAutonomous robotics