Affordance-Based Grasping Point Detection Using Graph Convolutional Networks for Industrial Bin-Picking Applications.

Area of Science:

• Robotics
• Computer Vision
• Machine Learning

Background:

• Grasping point detection is crucial for robotics and computer vision.
• Deep learning methods show promise for predicting grasping points, especially for object affordances in bin-picking.
• Existing methods often rely on RGB/RGB-D images, with limited clarity on 3D spatial information utilization.

Purpose of the Study:

• To adapt and apply Deep Graph Convolutional Networks (GCNs) for predicting object affordances using n-dimensional point clouds.
• To investigate the performance of GCNs for grasping point detection in an industrial bin-picking context for both suction and gripper end effectors.
• To develop a novel bin-picking oriented data preprocessing pipeline.

Main Methods:

• Adapted a Deep Graph Convolutional Network (GCN) model to process n-dimensional point clouds.
• Developed a specialized data preprocessing pipeline tailored for bin-picking applications.
• Created and utilized a high-accuracy RGB-D/3D dataset for model training.

Main Results:

• The GCN-based method achieved improved top-1 precision scores compared to a 2D Fully Convolutional Network (FCN).
• Achieved a 1.8% improvement in top-1 precision for suction grasping.
• Achieved a 1.7% improvement in top-1 precision for gripper grasping.

Conclusions:

• GCNs offer a performance boost for predicting object affordances in point clouds for robotic grasping.
• This research presents the first application of GCNs for predicting affordances for suction and gripper end effectors in industrial bin-picking.
• The developed method and dataset provide a flexible and effective solution for bin-picking applications.