DGCM-Net: Dense Geometrical Correspondence Matching Network for Incremental Experience-Based Robotic Grasping.

Area of Science:

• Robotics
• Artificial Intelligence
• Computer Vision

Background:

• Robotic grasping of novel objects remains a challenge.
• Learning from experience can improve grasping reliability.
• Transferring learned grasp strategies to unseen objects is crucial for adaptability.

Purpose of the Study:

• To present a method for grasping novel objects by learning from experience.
• To introduce a network for transferring learned grasp experience to unseen objects.
• To enable robots to achieve more reliable grasping over time.

Main Methods:

• Developed the dense geometric correspondence matching network (DGCM-Net) for encoding object geometry.
• Utilized metric learning for feature space encoding, enabling nearest neighbor search for relevant experience.
• Reconstructed 3D-3D correspondences using normalized object coordinate space for grasp configuration transfer.

Main Results:

• Achieved an equivalent grasp success rate compared to baseline methods.
• Significantly improved baseline methods by fusing experience knowledge with their grasp strategies.
• Demonstrated successful transfer of grasps to new instances and improved success rates with increased experience in offline experiments.

Conclusions:

• The DGCM-Net effectively transfers learned grasp experience to novel objects.
• The approach enhances grasping reliability and success rates over time.
• Learned task-relevant grasps can prioritize configurations for functional object use.