| dc.description.abstract | There is no question that autonomous vehicles and robots are going to have a profound impact on the society in the near future. A deep understanding of the systems is required in-order to design safe autonomous systems that are deployed among people. This thesis proposes a nonlinear data-driven system identification techniques to improve the physics-based models coupled with a nonlinear MPC for path tracking for the Warthog and ASV. This thesis also proposes a novel approach for a longitudinal controller. Domain knowledge was extensively used to generate a feature set for the machine learning algorithms to learn. During the experiment, it was found that for the Warthog dataset, it needed at least 103 seconds worth of data to outperform the physics-based model. No conclusion can be made regarding the generality of the model. The learned models, for both the Warthog and ASV, returned the physics-based model with the parameters identified, suggesting that there is a slight delay in the system that was not accounted for given the nominal values. The nonlinear MPCs were validated using the Gazebo and CARLA simulators. The MPC for the Warthog followed the predefined path with an average error of 8 cm, however, it struggled with sliding while turning. The MPC for the ASVs were tested using three different scenarios; highway exits and entrances, highway driving, and city driving. Both the lateral and the controller tracked the generated path within the safety margin, which is the lane width. However, it failed the city driving test, exceeding the cross-track error of 3 m. | en |
