Bridging Reinforcement Learning and Iterative Learning Control: Autonomous Motion Learning for Unknown, Nonlinear...

Area of Science:

• Robotics
• Machine Learning
• Control Theory

Background:

• Autonomous robots often face challenges with unknown, nonlinear dynamics, hindering precise reference tracking.
• Existing control methods typically require detailed system knowledge or manual tuning, limiting their real-world applicability.
• Validation in physical experiments remains scarce for many advanced robotic control techniques.

Purpose of the Study:

• To develop a novel learning control scheme for robots that can autonomously adapt to unknown nonlinear dynamics.
• To eliminate the need for prior system state knowledge or complex feedback structures in robotic control.
• To create a "plug and play" solution for reference tracking problems in robotics.

Main Methods:

• A Gaussian Process (GP) model is employed to learn and approximate the robot's unknown dynamics online.
• The learned GP model is used to generate an optimized feedforward control input for each experimental trial.
• Algorithm parameters are automatically determined, enabling a hands-off, user-friendly approach.

Main Results:

• The proposed method demonstrated effective reference tracking in extensive simulations and real-world experiments.
• Performance was robust across various learning parameters and multiple motion tasks.
• A balancing robot successfully learned to track desired outputs rapidly using the "plug and play" method.

Conclusions:

• The model-agnostic, "plug and play" learning control scheme offers a significant advancement for robotic systems with unknown dynamics.
• This approach is highly adaptable and shows great potential for a wide range of reference tracking applications.
• The method's ease of use and demonstrated effectiveness pave the way for broader adoption in complex robotic systems.