Neural-Network-Based Iterative Learning Control for Multiple Tasks.

Area of Science:

• Robotics
• Control Systems Engineering
• Artificial Intelligence

Background:

• Iterative learning control (ILC) excels at trajectory tracking for repetitive tasks, even with nonlinear dynamics.
• Traditional ILC is limited to repetitive tasks, restricting its application to multiple, non-identical trajectories.
• Existing methods often require significant memory to store data for numerous tasks.

Purpose of the Study:

• To develop a novel neural-network-based ILC (NN-ILC) capable of effectively handling nonrepetitive tasks.
• To propose a method that generalizes ILC outputs for multiple tasks into a single functional representation.
• To reduce memory requirements and enhance the predictive capabilities of ILC for trajectory tracking.

Main Methods:

• A position-based ILC framework is designed to compensate for tracking errors.
• ILC outputs for multiple tasks are decomposed into linear and nonlinear components.
• Complementary neural networks, including general and switching networks, estimate the nonlinear portion.
• The linear and nonlinear parts are combined into a unified neural-network-based function.

Main Results:

• The NN-ILC successfully compresses ILC outputs for multiple tasks into a single function, significantly saving memory.
• The proposed method demonstrates strong generalizability, enabling prediction of ILC outputs for new tasks without additional learning.
• Experimental results on a robot arm validate the effectiveness of NN-ILC for multi-task trajectory tracking.

Conclusions:

• The NN-ILC effectively extends ILC to nonrepetitive tasks, offering memory savings and improved generalization.
• This approach accelerates the iterative learning process by enabling prediction of ILC outputs for diverse tasks.
• NN-ILC presents a powerful solution for complex trajectory tracking control in robotics and beyond.