Machine Learning methods are being developed and applied into three aspects of the project:

1.  Road damages detection and classification
2.  Localization
3.  Reinforcement learning for autonomous driving

Road damages detection and classification

The quality of subsurface structures is essential for the maintenance and prevention of road damages. To this end, neural networks algorithms for detection and classification of subsurface damages extracted from ground-penetrating radar (GPR). Deep neural networks (DNN) have been developed for preprocessing, material- and damage classification of big datasets in the cloud, as well as small datasets for in-car inference.

As an extension of this task, the neural networks were used to also detect surface cracks and classify them.

Localization

The accurate localization of autonomous ground vehicles has being pursued in the past years due to its importance. In this project, the localization is improved as the combination of complementary techniques, as sensor fusion and machine learning.

Reinforcement learning for autonomous driving

Novel hierarchical reinforcement learning algorithms are developed and applied in the simulations and to our experimental vehicle for autonomous driving. End-to-end learning for self-driving and path planning are in the scope of the project.