A Robust Visual Grasping Method for Robots in Cluttered and Stacked Scenes

Area of Science:

• Robotics
• Computer Vision
• Artificial Intelligence

Background:

• Vision-based robotic grasping accuracy degrades in complex, occluded environments.
• Traditional methods face limitations with open-loop
• segmentation-then-estimation
• approaches.

Purpose of the Study:

• To develop an iterative closed-loop optimization framework for enhanced robotic grasping pose estimation.
• To improve object manipulation and grasping reliability in challenging, unstructured environments.

Main Methods:

• Coupling Segment Anything Model (SAM) with FoundationPose in a closed-loop system.
• Implementing a mask correction mechanism using 3D rendered projections for pixel-level boundary refinement.
• Utilizing a multi-dimensional confidence assessment integrating 2D and 3D domains for pose reliability evaluation.

Main Results:

• Achieved an ADD-S recall rate of 91.7% in cluttered and stacked scenes.
• Reduced average translation and rotation errors to 3.5 mm and 2.1°, respectively.
• Demonstrated a 96.5% grasping success rate in 200 real-world trials.

Conclusions:

• The proposed closed-loop optimization framework significantly enhances the robustness and accuracy of robotic grasping pose estimation.
• The synergistic integration of SAM, iterative refinement, and multi-dimensional confidence assessment proves effective in unstructured environments.