I. Introduction
Modern robots are equipped with a wide variety of sensors and actuators for observing and interacting with the surrounding environment. This allows a robot to carry out applications such as 3D reconstruction, object recognition, grasping, and much more. To receive further clues about its surroundings, a robot must move its sensor to another location and obtain new sensor values [1], [2], [3]. Several factors constrain the acquisition of the next sensor view including the kinematics of the robot, the field of view and range of the sensor, and environmental obstructions such as occlusions [4] and obstacles [5]. The challenge of obtaining a series of updated sensor placements is known as the next best view (NBV) problem. Specifically, given a known sensor pose and an information value, the next placement of the sensor should ensure that the value obtained maximizes the ‘information gain’ across the full action space of the sensor, Fig. 1. Information gain is calculated between two subsequent views, and the full gain of a sequence of views is the sum of the individual consecutive gains.