Data-Driven Object Pose Estimation in a Practical Bin-Picking Application.

Area of Science:

• Robotics
• Computer Vision
• Artificial Intelligence

Background:

• Pose estimation for textureless industrial metallic parts is challenging due to reflective properties and varying lighting conditions.
• Conventional vision-based methods struggle with the appearance variability of metallic objects in semistructured bin-picking tasks.

Purpose of the Study:

• To develop a robust and cost-effective solution for 2D pose estimation of metallic parts in industrial bin picking.
• To address the limitations of existing methods by utilizing direct lighting and a data-driven approach.

Main Methods:

• A data-driven approach using convolutional neural networks (CNNs) was employed, eliminating the need for hard-coded object geometry.
• Direct lighting was integrated with a 2D camera mounted on the robot's gripper to exploit object reflectivity.
• The solution was adapted for industrial settings and validated through extensive real-world factory testing.

Main Results:

• The proposed method successfully estimates the pose of textureless metallic parts, overcoming challenges related to reflectivity and viewing direction.
• The system demonstrated effectiveness in a semistructured bin-picking task within an industrial environment.
• A semi-automatic data-gathering process was enabled, simplifying on-site application.

Conclusions:

• The developed CNN-based approach offers a practical and efficient solution for robotic bin picking of metallic components.
• The use of direct lighting and a 2D camera provides a cost-effective alternative to complex vision systems.
• The method's adaptability and successful industrial testing highlight its potential for manufacturing automation.