Regression Based Model for Autosteering of a Car with Delayed Steering Response

Abstract

In our work we consider the area of intelligent road vehicles, especially, the topic of automated vehicles. Our objective is to automatically develop - by applying a supervised learning approach - the steering of a realistically simulated car featuring a steering delay and rate limit of turning of front wheels. Due to the adopted physical constrains, the typically used steering mechanisms (based on servo-control model) result in a non-stable, oscillating behavior of the controlled car. The proposed approach of automated development of steering for such a realistically simulated car employs the perception-action relationship obtained from the sample runs of an ideal car (featuring instant steering response) steered by the servo-control model. Then, implying that offsetting the perceptions in the obtained relationship back in time to the value equal to the steering delay would mimic the behavior of the car featuring steering delays, we used the so modified perception-action relationship to train the proposed regression based model. The experimental results verify that the derived automated steering controls the car featuring steering delays of 100 ms and 200 ms in a much similar way as the servo-control steers an ideal car. Moreover, for delay of 400 ms the steering, developed via proposed approach provides a better quality of control than that obtained from one of the most versatile unsupervised machine learning approaches - genetic programming. approach - the steering of a realistically simulated car featuringa steering delay and rate limit of turning of front wheels. Dueto the adopted physical constrains, the typically used steeringmechanisms (based on servo-control model) result in a non-stable, oscillating behavior of the controlled car. The proposed approachof automated development of steering for such a realistically simulatedcar employs the perception-action relationship obtainedfrom the sample runs of an ideal car (featuring instant steeringresponse) steered by the servo-control model. Then, implyingthat offsetting the perceptions in the obtained relationship backin time to the value equal to the steering delay would mimicthe behavior of the car featuring steering delays, we used theso modified perception-action relationship to train the proposedregression based model. The experimental results verify that thederived automated steering controls the car featuring steeringdelays of 100 ms and 200 ms in a much similar way as theservo-control steers an ideal car. Moreover, for delay of 400ms the steering, developed via proposed approach provides abetter quality of control than that obtained from one of the mostversatile unsupervised machine learning approaches - geneticprogramming.