Trajectory-Based Skill Learning Using Generalized Cylinders.

Area of Science:

• Robotics
• Machine Learning
• Geometric Modeling

Background:

• Robotic skill acquisition from human demonstrations is crucial for versatile robot applications.
• Existing methods often struggle with generalization and adaptability to new situations.

Purpose of the Study:

• Introduce Trajectory Learning using Generalized Cylinders (TLGC) as a novel approach for robot skill learning.
• To enable robots to learn, reproduce, and generalize skills from human demonstrations effectively.

Main Methods:

• TLGC models demonstrated skills using Generalized Cylinders, a geometric representation defining a skill's demonstration space.
• The approach extracts implicit skill characteristics and boundaries.
• It supports generating multiple skill reproductions and generalizing skills via trajectory editing.

Main Results:

• The geometric representation allows for skill generalization to unforeseen situations.
• TLGC facilitates obstacle avoidance and interactive human refinement through kinesthetic correction.
• Validation was performed using Jaco and Sawyer robotic arms in real-world experiments.

Conclusions:

• TLGC offers a robust framework for robot skill learning from demonstrations.
• The use of Generalized Cylinders provides a powerful geometric representation for enhanced skill generalization and adaptability.
• This method paves the way for more intuitive and flexible human-robot interaction in skill transfer.