Sensor Planning for 3D Visual Search with Task Constraints

Abstract

Visual search is a fundamental problem in autonomous robotics. Traditionally, visual search is formulatedas an optimization problem in which the sequence of actions ischosen based on immediate efficiency. In this paper we examinethe effects of the task constraint in the form of maximumallowable cost on action selection in search. We propose threealgorithms, namely Greedy Search with Constraint (GSC),Extended Greedy Search (EGS) and Dynamic Look AheadSearch (DLAS), to investigate which algorithm, whether locallyor globally, has the most efficient performance under variousconditions with a predefined task constraint. We examine ourmethods in environments of various sizes and configurationswith three cost constraints including time, energy consumptionand the distance travelled by the robot. Through extensiveexperiments on a mobile robot, we show that the environmentcharacteristics as well as the type of constraint applied canalter the performance of the methods significantly. We alsoshow that GSC algorithm, which relies on visual clues in anenvironment to optimize search, achieves the best and mostefficient performance in comparison to EGS and DLAS.