Grasping learning, optimization, and knowledge transfer in the robotics field.

Area of Science:

• Robotics
• Artificial Intelligence
• Computer Vision

Background:

• Service robotics demands intelligent manipulation for human interaction.
• Handling deformable objects with compliant robotic hands presents grasping control challenges.
• Existing simulation-based learning methods are ineffective for complex grasping dynamics.

Purpose of the Study:

• To optimize grasping strategies for unforeseen deformable objects using a compliant, under-actuated, sensorless robotic hand.
• To develop a sample-efficient learning approach for robotic manipulation tasks.
• To enhance grasping capabilities through transfer learning.

Main Methods:

• A Bayesian optimization approach was employed for grasping strategy optimization.
• RGB images and partial point cloud data were used for object grasping.
• Transfer learning was integrated to leverage acquired knowledge for new objects.
• A PAL Robotics TIAGo platform was utilized for experimental validation.

Main Results:

• The proposed Bayesian approach demonstrated superior sampling efficiency compared to grid sampling.
• The optimized grasping strategy achieved an 88% success rate on diverse objects.
• Transfer learning enabled effective generalization to partially new objects and containers.

Conclusions:

• Bayesian optimization with transfer learning is an effective and sample-efficient method for grasping deformable objects.
• The approach significantly improves robotic manipulation capabilities in real-world scenarios.
• This research advances the development of versatile service robots capable of complex tasks.