Title: Learning control for autonomous vehicle guidance
Other Titles: Lerende controle voor autonome voertuig navigatie
Authors: Kayacan, Erkan; R0268785
Issue Date: 10-Dec-2014
Abstract: In agricultural field work, machines must be accurately navigated to achieve an optimal result. The entire field should be covered with minimal overlap during tillage, fertilizing and spraying. The rows must be nicely parallel nicely for sowing and planting and evenly distributed so that the weed hoes can be easily driven between them. The fact that this is a difficult task, can be clearly seen from the considerable overlap and variation in plant distances observed in the field. Moreover, this limits the speed at which the tasks can be performed. Apart from navigating the machine, the operator must also supervise the work performed by the machine. Switching between paying attention to the steering and paying attention to the machine control results in an increase in the deviation from the optimal path in practice. To alleviate the task of the operator and allow him to concentrate on the quality of work performed, systems were developed for driver assistance and semi-autonomous control. These systems initially worked primarily based on local positioning. Thanks to the developments in the field of RTK GNSS systems with cm accuracy, this is nowadays mainly done through global positioning. The current commercial systems use simple controllers to minimize the deviation from the target path by adjusting the steering angle. These systems work well for the following straight lines under uniform soil conditions with a constant speed. However, when the soil conditions or speed change, the controllers must be tuned again. Furthermore, they use independent controllers for the absolute steering of the tractor and the relative steering of the trailer. Since both controllers will exhibit selfish behavior, this often leads to a sub-optimal result, especially for curved target paths in which the steering action of the tractor works against that of the trailer. Advanced control methods are needed to achieve higher control accuracy for the trailer for both straight and curved target paths under varying soil conditions and to make tractor control subordinate to the trailer control. Therefore different advanced model predictive control structures were elaborated in this PhD research and implemented on an autonomous tractor-trailer system to investigate its potential in practice. Firstly, an adaptive kinematic model and a yaw dynamics model of the tractor-trailer system have been derived from first principles. Then, the longitudinal dynamics and the dynamics of the steering mechanisms have been identified. The overall resulting model can serve as a benchmark system for evaluating model based controller and estimator designs. When model-based control structures have to deal with uncertain and varying process conditions, it is inevitable to use adaptive models. Real-time estimators allow to make these model adaptations through online parameter estimation. In this study, nonlinear moving horizon estimation method has been chosen as a state and parameter estimation algorithm because it considers the state and parameter estimation within the same problem and allows to incorporate constraints both on states and parameters. Secondly, a nonlinear model predictive controller has been designed in which a full model of tractor-trailer system was used including all interactions. This centralized controller managed to let the system follow the target trajectory with a high accuracy. However, it requires a relatively high computational cost due to the complexity of the optimization problem up to 7.2 ms. To reduce the computational cost a fully decentralized nonlinear model predictive controller was also designed and implemented. However, as these decentralized controllers ignore the interactions between the subsystems, they affect each other's performance negatively and may destabilize each other. In order to make the system robust against the differences between the real system and the sub-models, a tube-based approach was applied for the interactions between the two subsystems. Thanks to dividing the optimization problem into two smaller problems, the computation time could be reduced by a factor of 5. However, the trajectory error increased by 20 to 50% by ignoring the interactions. As both the centralized and decentralized approaches had their merits and limitations, a compromise was sought in the form of a distributed nonlinear model predictive controller in which the interactions are partially taken into account. These distributed controllers were between the centralized and decentralized controllers in terms of both computation time and tracking error. Finally, as an alternative approach to reduce the required computation time, two linear model predictive controllers have been designed: a linear time-invariant and a linear time-varying. Only the linear time-invariant model predictive controller was capable of driving the tractor-trailer system on any desired trajectory with an acceptable accuracy. While the computation time decreased to $0.7$ ms which is a factor of 10 shorter than the centralized nonlinear model predictive control, the trajectory error increased by 50 to 100%. Nonlinear model predictive controllers and nonlinear moving horizon estimators are capable of tracking straight and curved lines accurately. Thanks to the input-state linearization, successful results have been reported for the linear time-invariant model predictive controller. However, the linear time-varying model predictive controller obtained by linearizing around the reference trajectory is not able to follow curvilinear trajectories due to the fact that the trajectory tracking is a nonlinear control problem from its nature. As future studies, nonlinear control and estimation methods are more than welcome. In addition to efficient numerical methods, decentralized and distributed nonlinear moving horizon estimators can be studied to decrease computation time in practice. Robot sensing is also welcome to make agricultural robots being aware of environment and making decisions without human intervention.
ISBN: 978-90-8826-384-2
Publication status: published
KU Leuven publication type: TH
Appears in Collections:Division of Mechatronics, Biostatistics and Sensors (MeBioS)

Files in This Item:
File Status SizeFormat
thesis.pdf Published 2670KbAdobe PDFView/Open Request a copy

These files are only available to some KU Leuven Association staff members


All items in Lirias are protected by copyright, with all rights reserved.