Viewpoint Planning Based on Uncertainty Maps Created from the Generative Query Network

Abstract

Most current research on robotic tasks such as grasping uses the single fixed viewpoint positioned to oversee the environment from above. However, there are cases when a single fixed viewpoint does not provide sufficient information to perform a designated robot task. Environments with high occlusion, or cases where objects have fragile components, it is difficult for robots to identify the task relevant viewpoint to achieve a task or goal. Herein, to reconstruct the world model with less effort, we propose a viewpoint planner method that uses uncertainty in scene representation from Generative Query Network(GQN) With this scene representation, we create an uncertainty map by calculating the pixel-wise variance of multiple predicted images for each query viewpoint. The results indicate that our solution is capable of taking a suitable view of an untrained object, therefore reducing the uncertainty. A suitable viewpoint is defined as one that improves prediction certainty by focusing an area in which the learning environment allows the highest uncertainty value. In a future study, we would like to propose a task specific viewpoint planner models for robotic tasks such as grasping objects with occlusion by extending this research.