Bandit-Based Online Candidate Selection for Adjustable Autonomy

Abstract

In many robot navigation scenarios, the robot is able to choose between some number of operating modes. One such scenario is when a robot must decide how to trade-off online between human and tele-operation control. When little prior knowledge about the performance of each operator is known, the robot must learn online to model their abilities and be able to take advantage of the strengths of each. We present a bandit-based online candidate selection algorithm that operates in this adjustable autonomy setting and makes choices to optimize overall navigational performance. We justify this technique through such a scenario on logged data and demonstrate how the same technique can be used to optimize the use of high-resolution overhead data when its availability is limited.