Deep Learning-Powered Embodied Navigation in Simulated Environments

This thesis focuses on embodied navigation using deep learning techniques in simulated environments. It proposes a self-supervised learning framework and a scene-wise data augmentation method to improve embodied navigation policies. Additionally, it explores advanced navigation skills by studying novel embodied navigation tasks, including the scenario oriented object navigation task and the cooperative indoor navigation task. This thesis extends our understanding of embodied navigation and contributes to the development of more intelligent navigation agents.