Deep Learning Based Control System for a Small Ground Vehicle

Abstract

The research and development of self-driving cars and unmanned vehicles capable of autonomous driving have been actively pursued, and a deep-learning neural network (DLNN) is being applied due to recent advances in computing power and numerical methods. In this paper, we investigate whether the DLNN can be applied as the nonlinear feedback controller to command a micro-ground vehicle (MGV) to follow the lateral reference position of the course coordinates. In order to generate supervised training data for the DLNN controller, a large number of first-order preview model simulation data starting from various initial conditions were computed. The utility of the proposed idea is evaluated by both the numerical simulation and experimental testing of the lane-keeping and the double lane change maneuver using the DLNN controller on the straight and curved sections in the oval-formed course. Results demonstrate the feasibility of commanding the MGV with the proposed nonlinear feedback controller based on the DLNN.