Robust Learning-Based Control for Uncertain Nonlinear Systems With Validation on a Soft Robot.

Area of Science:

• Robotics
• Control Theory
• Machine Learning

Background:

• Traditional control methods struggle with nonlinear and uncertain systems.
• Existing deep Koopman operator and reinforcement learning methods have limitations in robustness.

Purpose of the Study:

• To present a universal and robust control framework for uncertain nonlinear systems.
• To guarantee robust stability using a data-driven deep stochastic Koopman operator (DeSKO) model.

Main Methods:

• Developed a deep stochastic Koopman operator (DeSKO) model to learn system uncertainty by inferring observable distributions.
• Designed a robust learning control framework with integral action for model predictive control.
• Integrated inferred uncertainty distributions into a stabilizing closed-loop controller.

Main Results:

• The DeSKO framework demonstrated superior robustness and scalability compared to state-of-the-art deep Koopman operator and reinforcement learning controllers in simulations.
• The method successfully resisted previously unseen uncertainties, including external disturbances up to five times the maximum control input.
• On a soft robotic arm, the DeSKO framework outperformed model-based controllers and executed pick-and-place tasks without retraining.

Conclusions:

• The DeSKO approach offers a robust solution for controlling high-dimensional nonlinear systems with internal or external uncertainties.
• This framework simplifies high-level control and decision-making in robotics.
• It opens new avenues for managing uncertainty in complex dynamic systems within a learning framework.