GripperPose: A novel approach for large industrial grippers 5D pose estimation
Abstract
Object pose estimation remains a challenging task in computer vision, especially for facilitating robotic manipulation. In industrial scenarios involving large-scale grippers—such as those used in waste incineration, port operations, and steel smelting—traditional pose estimation methods often prove inadequate. Accurate estimation of the spatial operating pose of these grippers is essential for fault diagnosis and ensuring operational safety. However, existing pose estimation approaches are rarely applied in such environments, and directly transferring them often leads to domain adaptation issues due to complex backgrounds and the limited availability of depth information. Furthermore, representing the pose of rotationally symmetric grippers using conventional 6 degrees of freedom (DoF) introduces ambiguities, making precise estimation difficult. To address these challenges, this paper proposes GripperPose, a novel framework for 5D pose estimation specifically designed for large industrial grippers. The pose estimation task is redefined using a 5D representation, along with three new evaluation metrics tailored to the unique characteristics of rotationally symmetric objects. Additionally, we introduce Gripper10, a dedicated dataset for the 5D pose estimation of ten commonly used industrial grippers in waste-to-energy incineration plants. Experimental results demonstrate that GripperPose achieves strong robustness and adaptability in these complex industrial environments.