Robot multi-target high performance grasping detection based on random sub-path fusion.

Area of Science:

• Robotics
• Computer Vision
• Machine Learning

Background:

• Grasping objects with uncertain shape, attitude, scale, and stacking presents a significant challenge in robotics.
• Existing methods often struggle with the complexity and variability of multi-target object manipulation in unstructured environments.

Purpose of the Study:

• To develop a high-performance grasping network for pixel-level detection of multi-target objects with uncertain properties.
• To enhance the robustness and accuracy of robotic grasping systems in complex, real-world scenarios.

Main Methods:

• Proposes the random sub-path grasp fusion network (RSPFG-Net) utilizing an agile grasping representation (AGR) strategy.
• Introduces the Multi-Scale random sub-path fusion (MSRSPF) module to prevent overfitting and improve robustness.
• Integrates the MSRSPF module with the DeepLab v3 network for pixel-level grasping and multi-target grasp detection.

Main Results:

• RSPFG-Net achieved an average grasping detection accuracy of 97.85% on the Cornell, Jacquard, and NEU-MGD datasets.
• In real-world experiments, the robot demonstrated an average grasping success rate of 94.31% for multi-target objects.
• The proposed methods showed excellent performance and robustness in handling objects with uncertain attributes.

Conclusions:

• RSPFG-Net effectively addresses the challenges of multi-target grasping with uncertain object properties.
• The developed AGR strategy and MSRSPF module significantly contribute to improved grasping accuracy and system robustness.
• The study validates the practical applicability and high success rate of the proposed network in real-world robotic manipulation tasks.