LSTM-Enhanced TD3 and Behavior Cloning for UAV Trajectory Tracking Control.

Area of Science:

• Robotics
• Artificial Intelligence
• Control Systems

Background:

• Unmanned aerial vehicles (UAVs) struggle with precise trajectory tracking in dynamic environments due to nonlinear dynamics and external disturbances.
• Existing control methods often require accurate system models, which are difficult to obtain for complex UAV operations.

Purpose of the Study:

• To develop a novel bio-inspired deep reinforcement learning (DRL) algorithm for high-precision UAV trajectory tracking.
• To enhance UAV control policy learning without relying on an accurate system dynamics model.

Main Methods:

• Integration of behavior cloning (BC) and long short-term memory (LSTM) networks within the Twin Delayed Deep Deterministic Policy Gradient (TD3) algorithm (TD3-LSTM-BC).
• Utilizing LSTM for temporal pattern recognition to improve adaptability to trajectory variations and resilience to disturbances.
• Employing BC for pre-training the DRL policy with expert demonstrations to accelerate convergence.

Main Results:

• The proposed TD3-LSTM-BC method demonstrated superior robustness and tracking accuracy in simulation experiments.
• The LSTM module enhanced the policy network's ability to capture temporal state patterns, outperforming memoryless networks.
• Behavior cloning effectively initialized the policy, reducing exploration time and improving learning efficiency.

Conclusions:

• The TD3-LSTM-BC framework offers a robust and accurate control solution for autonomous UAVs in challenging environments.
• Bio-inspired deep reinforcement learning, combining imitation and adaptive optimization, shows significant potential for UAV control.
• The method successfully mimics natural learning processes, balancing innate guidance with experiential adaptation for improved performance.