Sequential Inspection Using Loitering

Abstract

A set of objects of interest is to be sequentially inspected by a Micro Aerial Vehicle (MAV) equipped with a camera. Upon arriving at an object of interest, an image of the object is sent to a human operator, who, upon inspecting the image, sends his feedback to the MAV. The feedback from the operator may consist of the pose angle of the object and whether he has seen any distinguishing features of the object. Upon receiving the feedback, the MAV uses this information to decide whether it should perform a secondary inspection of the object of interest or continue to the next object. A secondary inspection has a reward (or value or information gain) that is dependent on the operator’s feedback. There is an associated cost of reinspection and it depends on the delay of the operator’s feedback. It seems reasonable to let the MAV loiter for a while near the most recently inspected object of interest so that it expends a small amount of endurance from the reserve after receiving the feedback from the operator. The objective is to increase the information and hence, the total expected reward about the set of objects of interest. Since the endurance of the MAVs is limited, the loiter time near each object of interest must be carefully determined. This paper addresses the determination of the optimal loiter time through the use of Stochastic dynamic programming. Numerical results are presented that show the optimal loiter time is a function of the maximum expected operator delay.