Development and Comparison of Neural Network-based Trajectory Planners for End-to-end Autonomous Driving

Trajectory planning involves determining the optimal path and control signals for a vehicle to navigate through a given environment, considering the surrounding conditions and potential obstacles. This thesis presents a project conducted at VisLab, an Italian company renowned for its research in autonomous driving, aiming to develop a low-level planning system based on Deep Neural Networks (DNNs). The objective is to enable the vehicle to navigate by generating a sequence of control signals that represent the vehicle's intentions in the form of subsequent acceleration and curvature values. We implement three distinct architectures and trained through Imitation Learning (IL) on urban scenarios from the city of Parma. Our methods are compared using different metrics related to both imitation accuracy and driving behaviour. Integrating vehicle dynamics into the loss computation leads to an improvement for trajectory prediction, while deriving the loss function directly from the output actions proves to be the best choice for control. Therefore, we design a model that combines this two properties into a composed solution, with prediction and planning heads. Hidden features extracted from the prediction module are shared with the control head, improving its accuracy.