Feedback-Based Informative AUV Planning from Kriging Errors

This paper presents a methodology to find a feedback-based plan for AUVs to visit important locations to estimate spatiotemporal fields. A model based on Kriging method is created containing predictions and errors in prediction. These errors are then used to develop a reward function that gives a high reward to a location with large error prediction and penalizes the energy consumption is created. A feedback plan is found using value-iteration based Markov Decision Model (MDP). The feedback plan helps adjust the paths or trajectories of AUVs in the presence of modeling errors and other forms of uncertainty due to the unpredictability of future states. Moreover, if the AUV deviates from its path during a mission, there will not be any need to re-plan since the plan will be calculated over the entire free space. Data were used from a mission at Big Fisherman’s Cover, at Santa Catalina Island, California, USA. Results show that the Kriging method can be used as a prediction model for several properties of the water in an oceanic environment. Furthermore, an optimal navigation policy was presented in the 2-D marine environment that generates the best possible action in the simulated policy from any location of the environment to the goal location.