Robots are increasingly being deployed in our daily lives and across society, from autonomous delivery and warehouse logistics to service robotics and disaster response. Navigation is a crucial and foundational capability that enables robots to operate effectively in real-world environments. A key challenge to develop generalizable navigation policies that allow robots to navigate and interact with complex real-world environments. Real-world navigation remains highly challenging due to the challenges of operating in diverse and dynamic environments with limited sensing, complex human interactions, and varying terrain conditions. In this dissertation, we address four core challenges toward developing autonomous navigation capabilities in real-world environments: collision avoidance in dense crowds, place recognition in indoor environments, navigation over complex outdoor terrains, and long-range navigation.

To enable safe and agile robot movement in crowded indoor environments, we develop new learning-based algorithms, DenseCAvoid and CrowdSteer, which significantly improve the success rates in terms of collision avoidance over 50% in dense crowds, compared to previous approaches. For accurate localization in indoor spaces with varying lighting and sparse or repetitive perceptual features, we improve the accuracy of place recognition by 29.27% over state-of-the-art techniques using a two-stage recognition pipeline. For outdoor navigation, we address the challenge of traversing bumpy, slippery terrains and hills, where traditional methods often fail. Our approach, AdaptiveON, improves the stability and safety of robot navigation in such terrains. To extend navigation techniques to long distance (i.e., more than hundreds of meters) in outdoor environments, we address key subproblems including traversability analysis (MTG, ET-Former), directional reasoning (DTG, VL-TGS), and social adaptation (VL-TGS) in complex, large-scale outdoor environments. We analyze both geometric (MTG) and semantic (ET-Former) information from robot perception to estimate environmental traversability. We formulate directional reasoning as the imitation of human-like decisions, using diffusion models to mimic human behaviors across diverse scenarios (DTG). We further improve social navigation by interpreting pedestrian intentions, group formations, and crosswalks using vision-language models (VL-TGS). To address the lack of long-range navigation datasets, we present GND, a large-scale dataset that spans 10 university campuses, with 668 minutes of data and global traversability maps. Finally, we integrate these techniques into a unified approach, MOSU, to achieve long-range outdoor navigation.