Object pose estimation in industrial environments using a synthetic data generation pipeline

Abstract

The handling of objects is a crucial robotic skill for the automation of the production industry. The trend to use machine learning to estimate the 6D pose of objects is driven by higher robustness and faster processing times. Machine-learning based 6D pose estimation algorithms are available with varying estimation performance, robustness and flexibility. Suitable algorithms have to be selected based on use-case specific production requirements. A concept to evaluate these algorithms is presented. The generation of synthetic data based on the production requirements is proposed, followed by an evaluation of the algorithms to assess the generalization performance from generic benchmark datasets to custom industrial datasets. The overall pipeline is presented, realized and discussed.