Control of autonomous multibody vehicles using artificial intelligence

Abstract

The field of autonomous driving has been evolving rapidly within the last few years and a lot of research has been dedicated towards the control of autonomous vehicles, especially car-like ones. Due to the recent successes of artificial intelligence techniques, even more complex problems can be solved, such as the control of autonomous multibody vehicles. Multibody vehicles can accomplish transportation tasks in a faster and cheaper way compared to multiple individual mobile vehicles or robots. But even for a human, driving a truck-trailer is a challenging task. This is because of the complex structure of the vehicle and the maneuvers that it has to perform, such as reverse parking to a loading dock. In addition, the detailed technical solution for an autonomous truck is challenging and even though many single-domain solutions are available, e.g. for pathplanning, no holistic framework exists. Also, from the control point of view, designing such a controller is a high complexity problem, which makes it a widely used benchmark. In this thesis, a concept for a plurality of tasks is presented. In contrast to most of the existing literature, a holistic approach is developed which combines many stand-alone systems to one entire framework. The framework consists of a plurality of modules, such as modeling, pathplanning, training for neural networks, controlling, jack-knife avoidance, direction switching, simulation, visualization and testing. There are model-based and model-free control approaches and the system comprises various pathplanning methods and target types. It also accounts for noisy sensors and the simulation of whole environments. To achieve superior performance, several modules had to be developed, redesigned and interlinked with each other. A pathplanning module with multiple available methods optimizes the desired position by also providing an efficient implementation for trajectory following. Classical approaches, such as optimal control (LQR) and model predictive control (MPC) can safely control a truck with a given model. Machine learning based approaches, such as deep reinforcement learning, are designed, implemented, trained and tested successfully. Furthermore, the switching of the driving direction is enabled by continuous analysis of a cost function to avoid collisions and improve driving behavior. This thesis introduces a working system of all integrated modules. The system proposed can complete complex scenarios, including situations with buildings and partial trajectories. In thousands of simulations, the system using the LQR controller or the reinforcement learning agent had a success rate of >95 % in steering a truck with one trailer, even with added noise. For the development of autonomous vehicles, the implementation of AI at scale is important. This is why a digital twin of the truck-trailer is used to simulate the full system at a much higher speed than one can collect data in real life.