User Preference Elicitation for Unmanned Aircraft System Collaborative Search

Abstract: Employing unmanned aircraft systems (UAS) in search tasks requires either direct manual control on behalf of the pilot, or time-intensive way-point designation and route planning. In addition to flight planning, situational awareness about the airspace and operational environment must be maintained, which often requires the use of secondary observers to manage the increased workload. To address these gaps, we formulate the problem as a matter of eliciting the user’s preference regarding specific points of interest captured from aerial imagery that the user wants the UAS to visit. The algorithm is defined as a Partially Observable Markov Decision Process, which is tasked with querying the user to understand desirable locations. The algorithm interacts with the user by suggesting points, whose specifics are captured using a combination of hand-crafted and neural network learned features from aerial imagery. Our methods are validated through Monte Carlo testing on real world imagery and show improved performance over standard greedy algorithms.