Multidimensional Swarm Flight Approach For Chasing Unauthorized UAVs Leveraging Asynchronous Deep Learning

TL;DR

The paper addresses unauthorized UAV threats by transforming passive tracking into active interception through a multidimensional swarm guided by asynchronous DRL. A controller coordinates six tracking UAVs in 3D space using RSSI measurements, with four policy/value networks per axis operating under the A3C framework; channel dynamics are modeled with $ oldsymbol{eta}(d) = P_{ ext{tx}} - L(d) + oldsymbol{\xi}$ and $L(d) = L_0 + 10n\log_{10}(d/d_0)$, incorporating Rician fading and a uniform spatial correlation matrix $oldsymbol{C}$ with ρ and Doppler $f_d = (v/c) f_c$. The method defines action spaces $ oldsymbol{\mathcal{A}}^d = \{1,2,3,4,5 ight\,text{m}\}$ and $ oldsymbol{\mathcal{A}}^{x,y,z}=\{-4,-2,0,2,4 ight\,text{m}\}$, with axis-specific rewards and a four-network asynchronous training regime that optimizes policy and value estimates while promoting exploration via an entropy bonus. Experimental results demonstrate convergence of training rewards and robust interception performance under varying sampling frequencies $F$ and spatial correlations ρ, including analyses of movement counts and total tracking time, indicating practical effectiveness for real-time anti-UAV operations. The work integrates swarm coordination with asynchronous learning to deliver proactive interception capabilities, potentially enhancing airspace safety in diverse channel conditions and enabling extensions to multi-sensor fusion.

Abstract

This paper introduces a novel unmanned aerial vehicles (UAV) chasing system designed to track and chase unauthorized UAVs, significantly enhancing their neutralization effectiveness.

Figures (8)

• Figure 1: Proposed system based on swarm flight using deep reinforcement learning for chasing illegal UAVs.
• Figure 2: The architecture of the proposed neural network.
• Figure 3: Training architecture using shared learning and asynchronous update for the proposed model
• Figure 4: Examples of small scale fading of six channels between each tracking UAV and target UAV when $f_d=1$ Hz.
• Figure 5: Average training reward when $F=100$ Hz and $\rho=0.5$.