SELECTIVE ACQUISITION OF OPERATOR KNOWLEDGE FOR SAFER SEMI-AUTONOMOUS ROBOT OPERATIONS

We have developed techniques for a robot to compute its expected myopic gain in performance from asking its operator specific questions, such as questions about how risky a particular movement action is around pedestrians. Coupled with a model of the operator’s costs for responding to inquiries, these techniques form the core of a new algorithm that iteratively allows the robot to decide what questions are in expectation most valuable to ask the operator and whether their value justifies potentially interrupting the operator. We have performed experiments in simple simulated robotic domains that illustrate the effectiveness of our approach.