Robotic Visual Data Acquisition System Using Automatic Object Rearrangement and Annotation

Abstract

Training generalized robotic planning and decision-making models require a substantial amount of training data to enhance the model's performance, particularly in machine vision. Traditional methods often depend on manually creating scenes and annotating data, which can be very time-consuming. To address this issue, we propose a robot automatic data acquisition system comprised of robotic placement arrangement and automatic data collection. This innovative approach allows robots to autonomously collect visual data of objects from various perspectives. Conducted in a chemistry laboratory setting, we compared datasets constructed through manual acquisition and data augmentation. The results demonstrated that our method is competitive, highlighting its significant potential for constructing large-scale datasets.