Applying Uncertainty Maps created from Generative Query Network for a Viewpoint Planner

抄録

Most of current research in robot tasks such as grasping use single fixed view point which is positioned to oversee the learning environment from above. There are cases where a single fixed viewpoint does not provide sufficient information to perform the 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 the task goal. We propose a viewpoint planner which uses uncertainty in the scene representation from Generative Query Network(GQN). From this scene representation, we create an uncertainty map by calculating the pixel-wise variance of multiple predicted images for each query viewpoint. The results show our implementation is capable of locating a suitable view of an unlearned object. A suitable viewpoint is defined as the viewpoint which improves prediction certainty by observing an area where the learning environment shows the highest value for uncertainty.