Publication: Development of a learning based trajectory tracking controller for autonomous vehicles
Abstract
Araç teknolojilerindeki son gelişmeler ve ileri sürüş destek sistemlerinin hızla gelişmesiyle otonom araçlar gün yüzüne çıkmaktadır. Bir otonom araç sürücüsüz olarak bir noktadan başka bir noktaya gidebilme kabiliyetine sahip olan bir kara taşıtıdır. Bunu yaparken temel olarak; çevreyi algılama, karar verme, planlama ve planlanan yörünge ve hız takibi için kontrol yeteneklerine sahiptir. Bu tezde bir otonom yarış aracının kontrol yeteneğinde temel olan yörünge ve hız takibi problemini çözmek amacıyla, aracın dinamiği “bisiklet dinamiği” kullanılarak modellenmiş, bu dinamik model “çift şerit değiştirme” testlerinde doğrulanmıştır. Modellenen sistem doğrusal zaman değişen model öngörülü kontrol yöntemlerinde kullanılmıştır. Yörünge takibi için tasarlanan model öngörülü kontrolcüde ölçülen gürültü olarak belirlenen aracın sapma oranı, yapay sinir ağları kullanılarak modellenmiş olup böylelikle araç anlık olarak sapma oranını daima hesaplayabilir duruma gelmiştir. Sapma oranının belirlenmesi iki yöntem kullanılmış olup; ilkinde araç yarış pisti üzerinde defalarca sürülmüş, aracın anlık konumuna karşılık sahip olması gereken sapma oranı ileri beslemeli sinir ağları ile modellenmiştir. İkincil olarak ise, otonom yarış aracının yanal dinamiği ileri beslemeli sinir ağları ve tekrarlayan yapay sinir ağları ile birleştirilerek derin sinir ağları oluşturulmuş, elde edilen derin sinir ağı modeli aracın sapma oranını yüksek doğrulukta hesaplayabilmiştir. Son olarak, tasarlanan optimal kontrol algoritmaları yarış pisti üzerinde referans hız ve yörünge yarış profiline karşılık test edilmiş, kontrolcülerin performansları ve otonom yarış aracının her bir kontrolcüdeki tur zamanları karşılaştırılarak en iyi performansın öğrenme tabanlı model öngörülü kontrolcüler olduğu ortaya konmuştur.
With the recent advances in vehicle technologies and Advanced Driving Assistance Systems (ADAS), autonomous vehicles are emerging day by day. An autonomous vehicle is a land vehicle capable of driving from one point to another without a driver. While doing this, basically; it is capable of perception, decision making, planning and control for the planned trajectory. In this thesis, in order to solve the trajectory tracking problem, the dynamics of the vehicle was modeled using “bicycle dynamics” and validated in “double lane change” tests. This modeled system was used in linear time-varying model predictive control algorithms. The yaw rate of the vehicle, which is determined as the measured disturbance in the model predictive controller designed for trajectory tracking, was modeled using artificial neural networks, so that the vehicle can always calculate its yaw rate instantaneously. Two methods were used to determine the angular velocity. At first, the vehicle was driven many times along the race track, and the yaw rate that the vehicle should have in response to its instantaneous position was modeled with feedforward neural networks. Secondly, deep neural networks were created to estimate the lateral dynamics of the autonomous racing vehicle by combining feedforward neural networks and recurrent neural networks. Resulting deep neural network model was able to calculate the yaw rate of the vehicle with high accuracy. Finally; the designed optimal control algorithms were tested on the racetrack against the race trajectory. Performances of the controllers and the lap times of the autonomous racing vehicle in each controller were compared, and it was revealed that the best performance was on the learning based model predictive controllers.
With the recent advances in vehicle technologies and Advanced Driving Assistance Systems (ADAS), autonomous vehicles are emerging day by day. An autonomous vehicle is a land vehicle capable of driving from one point to another without a driver. While doing this, basically; it is capable of perception, decision making, planning and control for the planned trajectory. In this thesis, in order to solve the trajectory tracking problem, the dynamics of the vehicle was modeled using “bicycle dynamics” and validated in “double lane change” tests. This modeled system was used in linear time-varying model predictive control algorithms. The yaw rate of the vehicle, which is determined as the measured disturbance in the model predictive controller designed for trajectory tracking, was modeled using artificial neural networks, so that the vehicle can always calculate its yaw rate instantaneously. Two methods were used to determine the angular velocity. At first, the vehicle was driven many times along the race track, and the yaw rate that the vehicle should have in response to its instantaneous position was modeled with feedforward neural networks. Secondly, deep neural networks were created to estimate the lateral dynamics of the autonomous racing vehicle by combining feedforward neural networks and recurrent neural networks. Resulting deep neural network model was able to calculate the yaw rate of the vehicle with high accuracy. Finally; the designed optimal control algorithms were tested on the racetrack against the race trajectory. Performances of the controllers and the lap times of the autonomous racing vehicle in each controller were compared, and it was revealed that the best performance was on the learning based model predictive controllers.