Counter a Drone in a Complex Neighborhood Area by Deep Reinforcement Learning

Area of Science:

• Robotics and Autonomous Systems
• Artificial Intelligence
• Aerospace Engineering

Background:

• Counter-drone technology is rapidly advancing with artificial intelligence (AI).
• AI-powered systems offer enhanced accuracy and efficiency in drone threat mitigation compared to traditional methods.
• Autonomous systems are crucial for real-time threat detection and engagement.

Purpose of the Study:

• To propose a deep reinforcement learning (DRL) architecture for an autonomous counter-drone system.
• To train a learning drone to detect and intercept a target drone within a complex suburban environment.
• To evaluate the effectiveness of transfer learning in improving DRL agent performance and reducing training crashes.

Main Methods:

• A DRL architecture was developed for a learning drone equipped with a front camera for depth image capture.
• The state space included depth images and scalar parameters (velocities, distances, angles).
• Transfer learning was applied using pre-trained model weights to accelerate training and improve performance.

Main Results:

• The DRL agent successfully learned to detect and avoid both stationary and moving obstacles.
• Transfer learning demonstrated a significant improvement in initial training rewards (approx. 35 more).
• Transfer learning reduced training crashes by 65% when ground obstacles were included.

Conclusions:

• AI-driven counter-drone technology, particularly using DRL, offers a promising solution for enhanced aerial security.
• Autonomous drone systems can effectively neutralize threats while navigating complex environments.
• Transfer learning is a valuable technique for optimizing the training of autonomous counter-drone agents.