In context
In robotic grasping and assembly, accurate 6D pose estimation of industrial parts is critical, especially in cluttered scenes with stacking and occlusion. Traditional methods often struggle with occlusions, while deep learning approaches may lack precision. This research, published in 2023, addresses these challenges by combining neural networks with optimization algorithms.
What was reported
Researchers from Shanghai University and the Chinese Academy of Sciences proposed a monocular pose estimation method that integrates a multi-module neural network with a genetic algorithm. The method first uses YOLOv4-tiny for object detection and PVNet for segmentation and center prediction. An L-shaped boundary is constructed to define the visible region of the part, reducing interference from occluded areas.
The approach extracts edge information within the segmented region to generate a chamfer distance function delimited by direction angles. A matching function combines shape and edge matching, and a genetic algorithm performs large-scale search and small-scale optimization for fast convergence of the 6D pose.
Experiments using ArUco-coded parts showed that the method estimates pose in about 0.5 seconds. At a 420 mm observation distance, lateral translation error was within 1 mm, and average rotation angle error was within 2 degrees. Comparative tests demonstrated effectiveness in complex environments with occlusion.
Why it mattered
This work offers a practical solution for robotic bin picking and assembly, where accurate pose estimation under occlusion is essential. By combining the robustness of deep learning with the precision of model-based optimization, it improves the reliability and efficiency of industrial automation systems.
“The proposed method can effectively deal with the accurate estimation of industrial part pose in complex environment and improve the working efficiency of robot.”
Source: 《机器人》期刊 (robot.sia.cn) · Published 2023-03-11 · “一种多模块神经网络与遗传算法相结合的单目位姿估计方法”