무인 실험실 로봇 비전 시스템의 객체 검출을 위한 디지털 트윈 기반 합성 데이터 생성 및 도메인 적응

초록

This study presents a dataset construction and data augmentation framework for enhancing the object detection performance in robot vision systems deployed on automated or autonomous robots in unmanned laboratory environments. In such settings, reliable object detection is critical for monitoring experimental equipment, avoiding unexpected obstacles, and ensuring safe and efficient laboratory operations. However, acquiring large-scale annotated datasets in real-world laboratory environments is expensive and time consuming. To address this limitation, a three-dimensional (3D) digital twin of the laboratory environment was constructed and virtual image data were collected by synchronizing the robot and camera parameters between the digital twin and real world. A Cycle generative adversarial network (CycleGAN) was then employed to generate domain-varied synthetic images by learning style transformations from virtual, real, and external domain images. In addition, real target objects were composited into diverse synthetic environments to further enrich the dataset diversity. Object detection models trained on the proposed synthetic datasets demonstrated improved domain adaptation and superior detection performance in real- world laboratory scenes compared to models trained solely on virtual data. The proposed approach provides a cost-effective solution for training robust robot vision models while significantly reducing the need for extensive real-world data collection, thereby contributing to the advancement of autonomous robot vision systems in unmanned laboratory environments.

키워드