Pose Estimation of Industrial Textureless Objects via Fusion of Simulated and Real RGB-D Data

Accurate pose estimation of textureless objects is crucial for robotic automation tasks, such as industrial bin-picking, where objects are randomly piled, heavily occluded, and challenging to detect. Traditional methods relying solely on high-quality real data face limitations of high cost and low efficiency. In this paper, we propose a novel technique that fuses low-quality real RGB-D data from consumer-grade cameras and high-quality simulated RGB-D data generated using domain randomization. Our approach significantly improves the robustness and accuracy of pose estimation algorithms. Experimental results demonstrate superior performance in both simple and cluttered industrial scenarios, confirming the effectiveness and practicality of our approach.