Counter a Drone in a Complex Neighborhood Area by Deep Reinforcement Learning is an open-access MDPI work by Ender Çetin, Cristina Barrado, and Enric Pastor.

Counter-drone technology utilizing artificial intelligence (AI) is a rapidly developing field. Given recent advances in AI, counter-drone systems equipped with AI can be extremely accurate and efficient in combating drone threats. The engagement time with targets can be significantly reduced compared to human-intervention methods, such as using a machine gun to bring down a malicious drone. Additionally, AI can identify and classify targets with high precision, minimizing the risk of false interdiction.

AI-enhanced counter-drone technology will provide significant advantages in addressing drone threats, contributing to safer and more secure skies. In this study, the authors propose a deep reinforcement learning (DRL) architecture to counter a drone using another drone, referred to as the learning drone, which autonomously navigates a suburban neighborhood environment while avoiding obstacles.

The simulated environment includes stationary obstacles such as trees, cables, parked cars, and houses. Additionally, a non-malicious third drone acts as a moving obstacle within the environment. The learning drone is trained to detect stationary and moving obstacles and to counter and capture the target drone without crashing into any obstacles in the neighborhood.

Equipped with a front camera, the learning drone continuously captures depth images, which form part of the state used in the DRL architecture. The state also includes scalar parameters such as velocities, distances to the target, distances to defined geofences, track, and elevation angles. These images and scalar data are processed by a neural network that integrates the two-state components into a unified flow.

The authors also tested transfer learning using the weights from a fully trained model. This approach yielded a significant jump-start, achieving higher mean rewards (close to 35 more) at the beginning of training. Transfer learning also reduced crashes during training, with a 65% decrease in crash episodes when all ground obstacles were included.

Publication Date– April 2020

Counter a Drone in a Complex Neighborhood Area by Deep Reinforcement Learning contains the following major sections:

  • Introduction
  • Methods
  • DRL Model Definition
  • Training Analysis & Results
  • Further Model Results and Discussion
  • Conclusions

Post Image Credit- envatoelements by Tatiana_Mara