Deep Reinforcement Learning-Based End-to-End Control for UAV Dynamic Target Tracking.

Area of Science:

• Robotics and Control Systems
• Artificial Intelligence
• Computer Vision

Background:

• Unmanned aerial vehicles (UAVs) face challenges in dynamic target tracking due to motion uncertainty, limited onboard camera perception, and control constraints.
• Traditional modular control paradigms for UAVs are complex and may not adapt well to unpredictable target movements.

Purpose of the Study:

• To develop a novel deep reinforcement learning (DRL)-based end-to-end control method for UAV dynamic target tracking.
• To simplify the traditional modular control framework by directly using onboard camera images.

Main Methods:

• A DRL framework was established, utilizing onboard camera imagery for direct control command generation.
• A neural network architecture, specific reward functions, and a soft actor-critic (SAC)-based algorithm were designed to train the control policy.
• The policy network's output was directly translated into speed control commands for the UAV.

Main Results:

• Numerical simulations demonstrated the feasibility of the proposed DRL-based end-to-end control method.
• The developed framework successfully simplified the conventional modular control paradigm for UAVs.
• The UAV exhibited effective tracking of dynamic targets with rapid changes in speed and direction.

Conclusions:

• The proposed DRL-based end-to-end control method offers a simplified and effective solution for UAV dynamic target tracking.
• This approach enhances UAVs' ability to autonomously track targets in complex and dynamic environments.
• The study validates the potential of deep reinforcement learning in advancing autonomous aerial systems control.