Bin-Picking Solution for Industrial Robots Integrating a 2D Vision System

摘要

Bin picking is a problem wherein a robot integrating a camera takes objects out of a bin filled with unorganized. Most bin picking algorithms heavily rely on the segmentation of RGB-D data, which is costly in terms of system scalability as each robot requires a 3D camera. This paper proposes a low-budget solution for the bin picking problem, wherein the robotic arm only equips a single 2D camera. Based on the RGB dataset collected from the 2D camera, a deep learning-based model, known as Mask R-CNN, is utilized to conduct object classification, detection, and instance segmentation simultaneously. Also, the Multi-Task RefineNet framework is applied for training of monocular depth estimation models. According to the instance segmentation masks, the principal component analysis is performed to determine the 4-DOF poses (3D position and orientation) of every detected object instance. The experiment was conducted to evaluate the segmentation and monocular depth estimation models. The experimental results indicate significant accuracy of the developed AI models, as well as the high grasping accuracy rate of the bin-picking robot using the proposed algorithm.