Robot Learning System Based on Adaptive Neural Control and Dynamic Movement Primitives.

Area of Science:

• Robotics
• Artificial Intelligence
• Machine Learning

Background:

• Robot skill learning is crucial for automation.
• Existing methods often struggle with complex motion generation and real-time adaptation.
• Dynamic Movement Primitives (DMPs) offer a framework for modeling robot motion.

Purpose of the Study:

• To propose an enhanced robot skill learning system.
• To improve motion generation and trajectory tracking capabilities.
• To increase the robustness of robot learning in dynamic environments.

Main Methods:

• Utilized Dynamic Movement Primitives (DMPs) to model robot motion from demonstrations.
• Integrated Gaussian Mixture Models (GMM) and Gaussian Mixture Regression (GMR) to enhance DMP learning from multiple demonstrations.
• Developed a neural-network-based controller, incorporating a Radial Basis Function (RBF) neural network, for trajectory tracking and environmental compensation.

Main Results:

• The enhanced DMP model successfully extracted more skill features from multiple demonstrations.
• Generated motions were scalable in both space and time.
• The RBF neural network controller effectively compensated for dynamic environmental changes.
• Experiments on a Baxter robot validated the proposed system's effectiveness.

Conclusions:

• The proposed system significantly improves robot skill learning.
• The integration of DMPs, GMMs, and RBF networks enhances motion generation and trajectory tracking.
• The system demonstrates robustness and adaptability in dynamic settings.