Evaluating FoundationPose object registration accuracy for AR-guided industrial inspection

Abstract Augmented Reality (AR) can enhance industrial processes like manual inspections by increasing safety and efficiency. This isachieved by superimposing virtual guidance directly onto the physical environment, such as visual highlights, step-by-stepinstructions, or animated tool paths. The accurate alignment of virtual guidance depends directly on the precise registration ofreal-world components. This study evaluates the 6D pose estimation accuracy of the FoundationPose framework for registeringreal-world, complex industrial components (e.g., fittings, flanges) from a single RGB-D image without object-specific retraining.Our evaluation demonstrates robust performance that meets the requirements for AR assisted inspection, achieving mediantranslational deviations between 0.6 mm and 2.8 mm and a median rotational error of approximately 2.8° for plane-symmetriccomponents. While performance remained stable against partial occlusions and reflective surfaces, the results highlight thataccuracy is limited by key challenges, including performance degradation with symmetric and texture-poor objects and theneed for optimization for real-time mobile deployment. Despite these limitations, the high baseline accuracy identified in thisstudy validates the significant potential of FoundationPose for industrial AR applications that demand precise alignment ofvirtual overlays.