Design of a neural steering controller applied to car trajectory tracking

The car trajectory tracking is currently an investigated problem and numerous researchers are still working on it proposing different approches based on both iterative methods and deep learning strategies. Many solutions have been developed based on Model Predictive controller and Convolutional Neural Network supplying reliable results. In this work different models are designed to solve the regression problem related to the steering angle of the vehicle based on both lateral and angular errors with respect to the desired trajectory. The work has been developed using deep learning strategies as Feedforward Neural Network and its recurrent variations as Long-short Term memory network. The used database is developed by considering as the target variable the output of the MPC previously designed by the INVETT research group of University of Alcalà. The challenging hyper-parameter tuning is performed by using both automatic tools as Optuna and observations related to previous works due to the high request in terms of computational amount of time. The comparison between the different models is performed by leveraging mainly on the Root Mean Square Error in order to give a measurement of the reliability of the prediction also in the more challenging case. In the end, the obtained results will be discussed. The programming language adopted in the entire project is python and some specialized libraries as keras.