Single-Camera Multi-View 6DoF pose estimation for robotic grasping

Area of Science:

• Robotics
• Computer Vision
• Machine Learning

Background:

• Accurate 6DoF object pose estimation is crucial for robot grasping.
• Existing multi-view methods often require multiple cameras, increasing complexity and cost.
• Occlusion and gripper collision can degrade pose estimation accuracy.

Purpose of the Study:

• To develop a more accurate and cost-effective 6DoF pose estimation method for robot grasping.
• To address limitations of traditional multi-view approaches.
• To validate the proposed method with a new dataset and real-world experiments.

Main Methods:

• Introduced a Single-Camera Multi-View (SCMV) method using one monocular camera and robotic manipulator motion.
• Captured multi-view RGB image sequences with the SCMV approach.
• Created and utilized the T-LESS-GRASP-MV dataset for validation.

Main Results:

• The SCMV method achieved significantly higher 6DoF pose estimation accuracy compared to public algorithms.
• Quantitative experiments on a real robot manipulator confirmed the method's high accuracy.
• Demonstrated robustness through successful completion of an assembly task with an 80% success rate.

Conclusions:

• The SCMV method offers a simpler and more accurate solution for robot object pose estimation.
• The proposed approach enhances grasping capabilities in robotics.
• The T-LESS-GRASP-MV dataset provides a valuable resource for future research in this area.