Reinforcement Learning-Based Model-Free Controller for Feedback Stabilization of Robotic Systems.

Area of Science:

• Robotics
• Control Systems
• Machine Learning

Background:

• Model-free control is crucial for complex robotic systems.
• Existing methods often require accurate system models.
• Reinforcement learning offers a promising alternative for adaptive control.

Purpose of the Study:

• To develop a model-free state feedback controller for robotic applications using reinforcement learning.
• To adapt least-square temporal difference (LSTD) learning with linear quadratic regulator (LQR) as a baseline.
• To address complexities in learning algorithms using policy iteration techniques.

Main Methods:

• Adaptation of reinforcement learning (RL) functions with least-square temporal difference (LSTD) learning.
• Development of a model-free state feedback controller using LQR as a baseline.
• Application of classical least-square policy iteration for boundary conditions and parameter estimation.

Main Results:

• Successful implementation of the proposed controller for trajectory tracking and balancing control.
• Demonstrated robustness in real-time experiments with unmanned helicopters and balancer robots.
• Validation of the model-free approach for complex robotic control tasks.

Conclusions:

• The proposed RL-based model-free controller exhibits robust performance in challenging robotic applications.
• The integration of LSTD and policy iteration effectively manages learning complexities.
• This approach provides a viable solution for model-free control in robotics.