Vision-Based Flexible and Precise Automated Assembly with 3D Point Clouds

Abstract

Cartesian-based visual servoing with finite-time control employing deep-learning-based partial point cloud registration for automated robotic picking and assembly is proposed in this paper. A partial point cloud registration architecture including deep-learning-based and Iterative Closest Point (ICP) algorithm for rough and precise registration is developed to resolve the time-consuming matching problem. Specifically, the deep-learning-based algorithm employs a descriptor to downsample and extract rough corresponding points that works as input to convolutional neural networks to estimate transformation between two point clouds. Then, ICP approach is applied to refine the alignment between them. For visual servoing, two kinds of controllers including Cartesian-based visual servoing and finite-time visual servo control are developed to drive the end-effector of the manipulator to reach desired pose for robotic picking and assembly. In addition, a pure force controller is also synthesized and implemented to resolve robotic insertion based on force/torque sensing to improve flexibility in handling automated robotic manufacturing in the proposed system. Finally, the performance of the proposed approaches has been verified by experiments with an industrial manipulator in velocity servo mode. Based on the analysis and experimental results in this paper, it appears to have demonstrated potential contributions and industrial applications in intelligent and flexible robotics manufacturing based on deep learning.