Supervised and Imitation Learning in Autonomous Vehicle Driving

Autonomous vehicle driving systems face the challenge of providing safe, feasible and human-like driving policies. The traditional modular approach involves a search-based or optimization-based motion planning followed by a feedback model-based controller. This may prove to be inadequate due to model uncertainties, limited computation time and difficulties in incorporating personalized and natural behaviour. The more recent end-to-end approach aims at overcoming these issues by learning from real drivers' data a policy to map from sensor data to controls using deep learning. Although being attractive by its simplicity, it also shows some drawbacks such as sample inefficiency and difficulties in validation and interpretability. The thesis presents mid-to-mid approaches attempting to exploit the best of both worlds, combining machine learning-based and model-based control into supervised and imitation learning frameworks that mimic expert driving behaviour from demonstrations while guaranteeing safety. To do so, the learned driving policies are used as guidance for model-based feedback control. In order to obtain realistic demonstrations, the training data comes from high-fidelity simulations of vehicle dynamics and advanced algorithms such as Model Predictive Control. Neural networks trained with supervised learning are shown to be viable as trajectory planners and feedforward controllers in the domain of time-finite safety-critical maneuvers. Imitation learning, with online data augmentation, is rather employed for sequential planning of standard driving trajectories. For this purpose, a smooth spline-based motion planning represents the policy provided by a constrained neural network exploiting the convex hull property of B-splines to enhance safety and reduce training time.