System identification in Autonomous Racing using AI (MSc AI)
My MSc thesis was a project in collaboration with the Technology Innovation Institute (TII) autonomous car racing team (TII EuroRacing) under the supervision of Jonas Eschmann, Professor Eliseo Ferrante and Dr Claudio Zito.
Here’s the abstract of the final work
System identification plays a crucial role in modelling dynamical systems with parametric models. In the case of vehicle dynamics modelling there are some barriers to overcome. They include the challenges in collecting data from realistic behaviors with the real vehicle, the complexity of capturing the system dynamics in a receding future horizon and adapting the vehicle model to compensate for rapidly changing dynamics in the environment. In this work, we formulate the system identification of vehicle dynamics as a mathematical optimization problem. The parameter identification of the vehicle dynamics model focuses on minimizing the error in the vehicle velocities predictions. We experimentally validate that the optimization methods are able to identify suitable parameters and enable the vehicle dynamic model to approximate the vehicle position over time more accurately when compared to the baseline model. Besides, we propose a behavior adaptation module that aims at modifying the vehicle dynamics model in an online fashion. The behavior adaption module is intended to enable the vehicle dynamics model to capture the underlying dynamics of the vehicle more precisely when comparing to the offline optimized vehicle dynamics model. Through experimental analysis, we show that behavior adaptation module was not able to improve the vehicle dynamics predictions accuracy when compared to the offline optimized vehicle dynamics model. Furthermore, we reflect upon the potential causes for the shortcomings of this approach.
– Jesus, A. F. (2022). System identification of vehicle dynamics for Autonomous Racing [Unpublished master’s thesis]. Vrije Universiteit Amsterdam.
The thesis focused on tyre model parameter identification and online tyre model parameter adaptation based on chronological vehicle dynamics simulation data. Motivated by the state-of-art in global multi-objective optimization, the parameter identification optimization problem was tackled with the Bees algorithm (D.T. Pham et al.) and compared to the least squares method. Additionally, the online parameter adaptation was formulated as a regression problem and tackled with a feedforward neural network.