Executive Summary : | Mobile robots, such as Autonomous/Unmanned Marine Vehicles, are widely used for ocean environment monitoring, security, search and rescue, and resource exploration. Optimal mission planning is crucial for minimizing costs and maximizing utility. Data collected during missions must be efficiently used on-board to re-plan paths, which is a critical link to imparting human-level intelligence to autonomous marine vehicles. On-board routing, or online re-planning, is a method that guides an autonomous agent to the most informative data and updates optimal paths on-board. The proposed learning method focuses on an autonomous agent with multiple sensors tasked with underwater sampling tasks in a stochastic and dynamic flow field. Marine robots are larger and more fidelity, making planning difficult. However, predictable environmental information helps in planning. An MDP framework for optimal planning in a spatio-temporal domain is developed, implemented on a GPU for offline planning. The current proposal aims to extend capabilities for on-board routing by deriving new update equations for the posterior MDP model from flow observations and implementing the same on GPUs for fast and scalable computation. Graph Neural Networks will be explored for on-board routing and implemented on GPUs. The software will be released as a package for end-users, making on-board routing an immediate requirement for practitioners in oceanography, the oil and gas industry, and naval operations. |