Autonomous navigation is a critical component of any exploratory mission. Implementing machine learning onboard spacecraft to improve optical navigation capabilities enables new and exciting mission opportunities. High-performance vision systems promise to improve autonomy relative to planetary bodies by being able to respond immediately to in-situ events, reduce reliance on NASA’s Deep Space Network for navigation, and eliminate the need to perform time-intensive Earth-based navigation tasks. Increased autonomy significantly boosts science discovery potential through higher mission efficiency at lower costs. However, neural networks act as black-box systems, which obfuscates interpretability during training or inference. Spacecraft navigation demands meeting strict and high accuracy requirements in unknown, unexplored, or potentially hazardous conditions. Limited autonomous navigation abilities in such environments limits exploration, especially if mission success is dependent on descent and landing. This research focuses on learning spatial awareness and exploring the strengths of different neural network architectures to improve upon the state of the art in optical navigation. This dissertation provides several contributions that improve autonomous optical navigation through the use of intelligent estimators: (1) onboard view synthesis to generate extra a priori data for navigation systems that do not yet have comprehensive knowledge of their surroundings, (2) onboard shape modeling without the need for human intervention, (3) high-accuracy navigation using multiscale and uncertainty-quantified super-resolution of remote sensing observations, and (4) state estimation through a fusion of machine learning and optimal filtering. These contributions take advantage of the intriguing and impressive aspects of machine learning while lowering risk through the use of trusted estimation algorithms. Thus, navigation becomes more robust and reliable.