(2) PUMA: Perception-aware UUV via Machine-learning-enabled Autonomy

In order for UUVs to be truly effective in the future maritime warfare, they have to be fully autonomous. They have to be self-sufficient, being able to make their own decisions and overcome problems without human or other external assistance. With the presence of set and drift as well as hardly perceivable obstacles and adversaries, navigating reliably undersea is a challenging task. To address these technical challenges, we propose to develop PUMA, a computational framework that enables a perception-aware UUV behavior in which a vehicle can autonomously take a course of actions that would improve its awareness of surroundings, so that the assigned mission can be effectively carried out. PUMA effectually employs machine learning to enable such intelligent behaviors. More specifically, PUMA incorporates two machine learning enabled components (LECs), namely, the Perception LEC and the Controller LEC. The Perception LEC will be pre-trained to detect a target object in a side-scan sonar image, and outputs the property of the detected object. The Controller LEC then takes that object property along with the measured vehicle/world state information, and outputs the desired vehicle velocity and direction based on the learned control law. The feasibility study will be conducted in Phase I.