Improved Adaptive-Reinforcement Learning Control for morphing unmanned air vehicles

Area of Science:

• Aerospace Engineering
• Control Systems
• Artificial Intelligence

Background:

• Unmanned air vehicle (UAV) morphing control presents challenges in achieving optimal shape change and precise trajectory tracking.
• Existing control methods, such as Actor-Critic algorithms and K-Nearest Neighbors (KNN), have limitations in accuracy and generalization.

Purpose of the Study:

• To develop an improved Adaptive-Reinforcement Learning Control (ARLC) methodology for UAV morphing control.
• To integrate reinforcement learning for optimal shape change policy with adaptive dynamic inversion for trajectory tracking.
• To enhance learning generalization from quantized to continuous state-action spaces.

Main Methods:

• Implemented an episodic unsupervised learning simulation using the Q-learning method to replace the Actor-Critic algorithm.
• Utilized Sequential Function Approximation (SFA), a Galerkin-based scattered data approximation, to replace the K-Nearest Neighbors (KNN) method.
• Integrated a reinforcement learning morphing control function with an adaptive dynamic inversion control trajectory tracking function.

Main Results:

• The improved ARLC methodology demonstrated smaller control errors compared to the KNN-based method.
• Enhanced learning of the optimal morphing shape was achieved.
• The SFA method effectively generalized learning from experienced states and actions to unexperienced continuous state-action spaces.

Conclusions:

• The proposed Q-learning and SFA-based ARLC method offers a more accurate and effective approach for UAV morphing control.
• This advanced control strategy improves both shape adaptation and trajectory tracking performance.
• The findings suggest significant potential for intelligent control in adaptive aerospace systems.