Weakly supervised metric learning for estimating autonomous vehicle terrain traversability cost

In this thesis, we address the challenge of learning the cost of traversing over arbitrary types of terrain using metric learning with weak labels. To address the challenge of manually labeling data, we construct our dataset by driving a HUSKY unmanned ground vehicle (UGV) throughout the UT Austin campus and leverage an anomaly detection method to generate pseudo labels to determine which areas in an image contain safe terrain to navigate on. This weakly labeled dataset is then fed into a neural network that is optimized to assign high rankings to untraversable terrain and low rankings to traversable terrain via a metric learning loss function. Our results demonstrate that the learned representations have the potential to be applied in fully autonomous driving tasks.