Learning-based fault-tolerant control for flexible joint robotic manipulators: An actor-critic framework with...

Area of Science:

• Robotics
• Control Systems
• Artificial Intelligence

Background:

• Flexible-joint robotic manipulators (FJRMs) offer enhanced dexterity but present control challenges due to nonlinear dynamics, coupling, disturbances, and actuator faults.
• Existing control methods often struggle with these complexities, particularly in fault-tolerant scenarios.

Purpose of the Study:

• To develop a robust fault-tolerant control strategy for FJRMs.
• To improve tracking accuracy and system stability in the presence of actuator faults and uncertainties.
• To leverage reinforcement learning for adaptive control policy optimization.

Main Methods:

• A Lyapunov-guided actor-critic reinforcement learning (RL) approach was implemented.
• A critic network was utilized to learn the long-run cost-to-go, considering tracking error, control effort, and fault impact.
• The critic's evaluation guided online actor policy improvement within a stabilizing Lyapunov control structure.

Main Results:

• The proposed strategy demonstrated improved tracking accuracy and robustness on a Baxter robotic platform under faulty conditions.
• Experimental results showed superior performance compared to PID and single neural-network controllers.
• Semi-global uniform ultimate boundedness of all closed-loop signals was theoretically established.

Conclusions:

• The Lyapunov-guided actor-critic RL strategy offers an effective solution for fault-tolerant control of FJRMs.
• This approach provides enhanced adaptability and resilience against uncertainties and actuator faults.
• The method represents a significant advancement over conventional control techniques for flexible robotic systems.