Design and Analysis of Data Synthesis Pipeline for Assembly Detection with Domain Randomization

We propose a framework that leverages product CAD models and domain randomization techniques to generate training data via simulators, enabling the training and fine-tuning of object detection networks entirely on synthetic data for real-world deployment. This framework is tailored for product assembly and intelligent manufacturing, where automation and time efficiency are critical. Focusing on assembly state detection and object pose estimation tasks, we also analyze the impact of multiple Domain Randomization (DR) variables and measure the time consumption across configurations. Experimental results demonstrate the feasibility and necessity of using mature simulators to address current challenges associated with the resource-intensive nature of data collection and annotation.