Human-robot planning and learning for marine data collection

We propose an integrated learning and planning framework that leverages knowledge from a human user along with prior information about the environment to generate trajectories for scientific data collection in marine environments. The proposed framework combines principles from probabilistic planning with nonparametric uncertainty modeling to refine trajectories for execution by autonomous vehicles. These trajectories are informed by a utility function learned from the human operator's implicit preferences using a modified coactive learning algorithm. The resulting techniques allow for user-specified trajectories to be modified for reduced risk of collision and increased reliability. We test our approach in two marine monitoring domains and show that the proposed framework mimics human-planned trajectories while also reducing the risk of operation. This work provides insight into the tools necessary for combining human input with vehicle navigation to provide persistent autonomy.