Learning Compliant Box-in-Box Insertion through Haptic-Based Robotic Teleoperation.

Area of Science:

• Robotics and Automation
• Logistics and Supply Chain Management
• Machine Learning for Robotics

Background:

• Automating compliant object insertion, like box-in-box tasks in logistics, is challenging due to difficulties in modeling object deformation.
• Learning from Demonstration (LfD) offers a viable approach for complex robotic tasks that are hard to model mathematically.

Purpose of the Study:

• To develop and validate an automated system for the box-in-box insertion task using Learning from Demonstration.
• To address the challenges of modeling object deformation and enabling precise insertion with position-controlled robots.

Main Methods:

• A master-slave teleoperated robot system was used for haptic demonstration of the insertion task.
• Gaussian Mixture Regression was employed for probabilistic trajectory learning from demonstrations.
• Barycentric interpolation was utilized for generalizing the insertion task and adapting to object position variations.

Main Results:

• The proposed framework successfully demonstrated automated box-in-box insertion by utilizing the natural compliance of the boxes.
• The method enabled precise insertion even with a position-controlled robot, showcasing adaptability.
• Experimental validation confirmed the generalizability and repeatability of the developed strategy.

Conclusions:

• Learning from Demonstration combined with probabilistic methods and interpolation provides an effective solution for complex robotic insertion tasks.
• The approach successfully leverages object compliance, overcoming limitations in modeling deformation for automated packaging applications.
• The validated method offers a robust and adaptable solution for real-world logistics and packaging automation.