Long-Term Autonomy for AUVs Operating Under Uncertainties in Dynamic Marine Environments

There has been significant interest in recent years in the utility and implementation of autonomous underwater and surface vehicles (AUVs and ASVs) for persistent surveillance of the ocean. Example studies include the dynamics of physical phenomena, e.g., ocean fronts, temperature and salinity profiles, and the onset of harmful algae blooms. For these studies, AUVs are presented with a complex planning and navigation problem to achieve autonomy lasting days and weeks under uncertainties while dealing with resource constraints. We address these issues by adopting motion, sensing, and environment uncertainties via a Partially Observable Markov Decision Process (POMDP) framework. We propose a methodology with a novel extension of POMDPs to incorporate spatiotemporally-varying ocean currents as energy and dynamic obstacles as environment uncertainty. Existing POMDP solutions such as the Cost-Constrained Partially Observable Monte-Carlo Planner (POMCP) do not account for energy efficiency. Therefore, we present a scalable Energy Cost-Constrained POMCP algorithm utilizing the predicted ocean dynamics that optimizes energy and environment costs along with goal-driven rewards. A theoretical analysis, along with simulation and real-world experiment results is presented to validate the proposed methodology.

I. Introduction

Marine robotic systems continue to increase their ability to operate independently for progressively longer periods. Existing systems have demonstrated robust, autonomous operations for multiple hours and even days. However, persistent (long-term) navigation capabilities will be critically important for future marine robots as they will be required to operate over periods of days to weeks. While current navigation and mapping algorithms can function over substantial spatial extents, it is currently unclear how to extend these to deal with human-scale spatial and temporal dimensions, as well as deal with the uncertainty of an ever-changing environment. As we look to extend our understanding of the Earth's changing environment, we require these marine robots and robotic systems to comprehend variability across large-scale spatiotemporal dimensions (50 km and days to weeks) while reacting to a locally dynamic and uncertain environment as illustrated in Fig. 1.