Autonomous mobile robots have revolutionized industries and workflows by enabling navigation and task execution in diverse environments without constant human intervention. The typical pipeline of an autonomous mobile robot begins with processing sensor inputs through perception modules to extract relevant information about the robot’s environment. This extracted information plays a crucial role in enabling the robot to make informed decisions and formulate plans. Furthermore, quantifying the uncertainty associated with these predictions is often highly desirable, as it enriches the robot’s decision-making process, leading to safer and more informed actions. Once the robot has devised its plans and decisions, they are then executed by the control module. The control module’s ability to handle unexpected situations and maintain the integrity of the robot’s actions is important for ensuring the overall success of its missions in real-world scenarios. This dissertation addresses various problems in the autonomous mobile robot pipeline. The initial focus of this dissertation centers on perception and the quantification of the associated uncertainty in the context of autonomous driving. This dissertation first delves into a perception problem concerning accurate shape estimation for vehicles using stereo images. Subsequently, it presents a novel point cloud completion algorithm that explicitly leverages the sensor signal captured over consecutive time from vehicle tracks, which results in improved quality of complete shape estimates for vehicles. Then, this dissertation presents a solution to improve image-based perception tasks for autonomous driving under challenging weather conditions by leveraging image-to-image translation networks to convert adverse weather images into benign ones, thereby facilitating improved perception tasks. This research proposes a new framework to train image-to-image translation networks using coarsely-aligned image pairs, leading to significantly more superior results compared to the conventional unpaired training methods typically employed. Then, this dissertation explores a solution for probabilistic uncertainty quantification in prediction models, demonstrated through an application in visual localization. In the final part of this dissertation, the focus transitions from land (autonomous driving) to underwater, addressing a challenging control problem for fish-inspired robots. This research aims to design a reinforcement learning controller that ensures robust navigation of the fish robot even in the presence of actuation failures.