Bridging Simulation and Reality: A 3D Clustering-Based Deep Learning Model for UAV-Based RF Source Localization

TL;DR

The paper tackles the sim-to-real gap in UAV-based RF source localization by introducing an Enhanced Two-Ray propagation model that accounts for UAV dynamics and leg shadowing, coupled with a 3D clustering based RealAdaptRNet trained entirely on simulated data. The Enhanced Two-Ray model improves RSS prediction by incorporating UAV orientation effects and body shadowing, while RealAdaptRNet employs 3D clustering to create compact, trajectory-robust inputs for localization. Validated on the AERPAW testbed, the approach demonstrates strong generalization across unseen trajectories and achieves competitive real-world accuracy with a substantial reduction in model size (about 33.5x fewer parameters) and computational cost. These results illustrate the viability of sim2real transfer for RF source localization and pave the way for efficient, scalable UAV deployments in real environments.

Abstract

Localization of radio frequency (RF) sources has critical applications, including search and rescue, jammer detection, and monitoring of hostile activities. Unmanned aerial vehicles (UAVs) offer significant advantages for RF source localization (RFSL) over terrestrial methods, leveraging autonomous 3D navigation and improved signal capture at higher altitudes. Recent advancements in deep learning (DL) have further enhanced localization accuracy, particularly for outdoor scenarios. DL models often face challenges in real-world performance, as they are typically trained on simulated datasets that fail to replicate real-world conditions fully. To address this, we first propose the Enhanced Two-Ray propagation model, reducing the simulation-to-reality gap by improving the accuracy of propagation environment modeling. For RFSL, we propose the 3D Cluster-Based RealAdaptRNet, a DL-based method leveraging 3D clustering-based feature extraction for robust localization. Experimental results demonstrate that the proposed Enhanced Two-Ray model provides superior accuracy in simulating real-world propagation scenarios compared to conventional free-space and two-ray models. Notably, the 3D Cluster-Based RealAdaptRNet, trained entirely on simulated datasets, achieves exceptional performance when validated in real-world environments using the AERPAW physical testbed, with an average localization error of 18.2 m. The proposed approach is computationally efficient, utilizing 33.5 times fewer parameters, and demonstrates strong generalization capabilities across diverse trajectories, making it highly suitable for real-world applications.

Figures (6)

• Figure 1: Secnario in AERPAW AFAR challenge Lake Wheeler testbed setup for RFSL. The blue region indicates the permissible UAV flight area, while the RF source can be located anywhere within the green-shaded region. The figure also depicts the fixed-way-points spiral trajectory designed to sweep the area and collect RSS measurements for localization. The upper-right corner displays the UAV used in the experiment.
• Figure 2: Illustration of the two-ray model, showing the direct and reflected signal paths between the transmitter (RF source) and receiver (UAV).
• Figure 3: DL-based pipeline for RFSL.
• Figure 4: (a) The received RSS values collected using the AERPAW testbed along the UAV trajectory for an RF source positioned at $(x, y) = (20, 184)$. Five clusters are considered, with black boundaries indicating the regions where points belong to their respective clusters. (b) The proposed RealAdaptRNet uses 1D convolutional layers denoted as Conv1D: $k, s, f$, where $k$ is the kernel size, $s$ is the stride, and $f$ is the number of filters. In ResNet Block A, the last convolutional layer always uses a stride of 1, while the other parameters depend on the placement within the architecture.
• Figure 5: CDF of absolute error (in dB) for the free-space, two-ray, and enhanced two-ray models compared to real-world RSS data.