Nowadays one of the most rapidly growing field in the industry is autonomous driving. Developing driverless cars requires a lot of state-of-the-art technology, including new sensors, high capacity computational units and the software stack running on them.
One module of the full stack software is usually responsible to follow the previously planned trajectory. The classical control methods are not able to deal with the continuously changing environment and the broad operational range thus more complex and robust control algorithms are required for safe driving.
During my work, I analyzed different approaches used to control car-like robots. I summarized the basic concept of control theory and the classical control methods – state space control with state feedback and PID control – and addressed their limitations. The next step was analyzing the control method used by Stanford University in 2005 DARPA Challenge. Moreover, I researched the optimal control methods, and optimization-based control methods, including the neural network based approaches.
I did my thesis in the Department of Automation and Applied Informatics in Budapest University of Technology and Economics. Using the framework developed by the MotionPlanning project, I implemented an extended version of Stanley-control and got to know of the method of designing optimization-based control with the help of ACADO framework. For the neural network based approaches, I used OpenAI baselines framework. In the end I compared my results with the already existing control.