View Planning Based on Inference in Virtual World for Object Pose Estimation

Abstract

It is important to estimate objects' poses accurately when a robot manipulates them. However, the estimation accuracy of a pose estimator would significantly decrease when the estimator was applied to the actual robot because of inadequate observation of the robot. This paper proposed efficient view planning based on an object's pose inference in a virtual world to improve estimation accuracy. The proposed method was evaluated in both simulation and real environments. The experimental results show that the proposed method reduces the estimation error and achieves high accuracy in grasping tasks.