BLUF
Bottom Line Up Front:
North Carolina State University researchers have completed empirical characterization of how 5G cellular networks perform when serving drones in real-world conditions—a critical gap in knowledge as the industry rushes to integrate UAVs into national airspace. Using custom Android measurement tools at the AERPAW testbed, they modeled signal strength and throughput across altitude, distance, and orientation, showing that machine learning models outperform classical free-space path loss predictions. This work provides the measurement-validated foundation operators and system designers need to deploy cellular-connected drones safely and reliably.
The Problem With Flying Users
The promise is enticing: a drone operating beyond visual line of sight, piloted remotely via 5G cellular connectivity, streaming high-definition video from disaster sites or agricultural fields—all using the same infrastructure that powers ground mobile networks. The reality, however, is far messier.
Cellular networks are optimized for ground users. Base station antennas tilt downward to serve pedestrians and vehicles at street level. When a drone climbs to 50 or 100 meters, it suddenly enters territory no terrestrial network designer anticipated: line-of-sight paths to distant towers, altitude-induced shadowing effects, Doppler-induced handover churn, and uplink interference that can reach base stations dozens of kilometers away.
Standardization bodies like 3GPP have begun adding drone-specific features to 5G New Radio (NR) and LTE—UAV flight path reporting, beamforming optimizations, remote identification (BRID) mechanisms, and detection-and-avoidance (DAA) support in Release 18. But standards alone cannot close what researchers call the air-to-ground modeling gap: the lack of empirical, measured data describing how real 5G networks behave when serving aerial user equipment (UEs) operating across varying altitudes, distances, and orientations.
"The existing infrastructure of cellular networks is primarily tailored to serve user equipment situated on or near ground level," notes a recent comprehensive survey in the ScienceDirect journal Computers & Electrical Communications. "The efficient integration of aerial user devices introduces considerable challenges."
Enter AERPAW—the Aerial Experimentation and Research Platform for Advanced Wireless.
AERPAW: A Testbed Comes of Age
Housed at North Carolina State University in Raleigh and funded by the NSF's $100 million Platforms for Advanced Wireless Research (PAWR) initiative, AERPAW began operations in 2019 as an experimental facility to study the convergence of 4G/5G technology and autonomous drones. In March 2024, the PAWR Project Office announced a dramatic Phase 2 expansion: four additional wireless towers, a total of eight unmanned aerial vehicles, and a flight field roughly three times the original size. AERPAW was also designated an Open Test and Integration Center (OTIC) by the O-RAN ALLIANCE, signaling its emergence as a critical infrastructure for open radio access network research.
The testbed infrastructure reflects what Rudra Dutta, co-PI and interim associate head of computer science at NC State, calls a "tall order": programmable radios via software-defined radios (SDRs), programmable drones with precise flight control, and a programmable network—all embedded in the real world, made remotely accessible to researchers nationwide, and all operating within FCC and FAA regulatory constraints.
Central to AERPAW's capabilities are Ericsson commercial 5G equipment operating in the C-band (3.4 GHz, 100 MHz bandwidth, 4×4 MIMO) and legacy 4G LTE in band 66 (1.7/2.1 GHz). Two Android phones are mounted on each test drone: a Samsung S21 running the custom-built PawPrints measurement application, and a Samsung S23+ running the proprietary Keysight Nemo Handy tool—a combination that offers both open-source accessibility and vendor-grade deep cellular KPI logging.
The Measurement Campaign: Real Data From the Sky
A collaboration between NC State faculty and graduate researchers executed three distinct measurement campaigns during 2024–2025, each flying methodical trajectories while logging 4G and 5G physical layer metrics and application-layer throughput. The experimental design was deliberately comprehensive:
- Campaign 1: A drone traced a polygon around the base station at 50 m altitude, logging 4G KPIs.
- Campaign 2: Horizontal sawtooth trajectories at increasing distances from the BS, at 30 m and 50 m altitudes, focusing on 5G.
- Campaign 3: Rectangular pattern sweeps across the BS in north-south direction, capturing handover dynamics.
The measured KPIs included reference signal received power (RSRP), reference signal received quality (RSRQ), channel quality index (CQI), signal-to-interference-and-noise ratio (SINR), packet data convergence protocol (PDCP) throughput, and physical layer channel rank—each a proxy for some aspect of link quality that operators and mission planners need to predict and manage.
The terrain at Lake Wheeler Field Labs, located south of Raleigh, exhibits rural characteristics: trees, open fields, and light vehicular traffic—representative of the environments where agricultural drones, infrastructure inspection UAVs, and emergency response aircraft will operate.
The Key Finding: Classical Models Fail; Machine Learning Succeeds
The researchers compared three modeling approaches: (1) Free Space Path Loss (FSPL), augmented with the base station's measured antenna radiation patterns; (2) polynomial curve-fitting models, degrees 2–9; and (3) machine learning regressors—random forest, XGBoost (XGB), and simple neural networks.
The results were stark.
Free Space Path Loss, the simplest model, achieved root-mean-square error (RMSE) of 4.79 dB for the Samsung S21 device and 4.25 dB for the S23—serviceable for ballpark estimates but inadequate for operational planning. The FSPL model's failures concentrated at two points: when the drone flew outside the base station's antenna main lobe, and when elevation angles exceeded ±12.5 degrees. In one flight trajectory, FSPL mean absolute error reached 7.24 dB, nearly double the campaign average.
Polynomial models showed improvement. A degree-5 polynomial fit based on logarithmic distance achieved RMSE of 3.59 dB (S21) and 2.93 dB (S23)—meaningful gains. But higher-degree polynomials (7–9) suffered from overfitting and erratic behavior outside the training domain, a classical trade-off between model simplicity and fidelity.
Random forest models emerged as champions. The random forest achieved RMSE of 2.89 dB (S21) and 2.23 dB (S23), with R² scores of 0.82 and 0.883, respectively. XGBoost performed similarly well (RMSE 2.99 dB and 2.46 dB). Simple neural networks trailed, achieving RMSE in the 2.5–3.8 dB range depending on architecture tuning.
The practical implication: a machine learning model trained on 3–5 hours of aerial measurements can predict 5G signal strength to within ±2 dB across a known operational area, compared to ±4.5 dB for classical theoretical models. For command-and-control links carrying safety-critical telemetry, that difference can determine whether a drone can reliably complete its mission or must abort.
KPIs Across Altitude: The 30-Meter Advantage
A particularly striking finding emerged from comparing 30 m and 50 m altitude flights on the same horizontal sawtooth trajectory:
- RSRP: 4.69 dBm higher at 30 m (75.96% of flight duration)
- SINR: 3.5 dB higher at 30 m (75.09% of flight duration)
- PDCP throughput: 62.14 Mbps higher at 30 m (81.41% of flight duration)
Channel rank—a measure of MIMO spatial multiplexing opportunity—showed minimal altitude dependence, but displayed a striking spatial pattern: rank-4 clusters appeared near the BS and at the edge of the main lobe, while rank-1 occurred farther away but still within the beam. A linear discriminant analysis (LDA) decision plane achieved 94% classification accuracy, correctly predicting rank distribution based on distance and angular orientation alone (Equation 4 in the paper).
This finding has immediate practical implications: operators can exploit high-rank conditions near the BS for throughput-sensitive payloads (video streaming, sensor data dumps) and switch to diversity-optimized transmission in low-rank regions to maximize reliability.
The Handover Challenge: When Drones Confuse the Network
One subplot emerges from the data: handover behavior. During a polygonal trajectory around a multi-sector Ericsson base station, the drone triggered multiple handovers between two physical cell identities (PCIs). The network exhibited a peculiar delay—even as RSRQ (reference signal received quality) dropped below the "poor" threshold of −15 dB, the device remained associated with the weakening cell for extended periods before handover triggered. This behavior, familiar to cellular engineers in terrestrial scenarios, may be pathological for UAVs: it creates windows of vulnerability where a drone loses the ability to receive reliable control commands precisely when it should be handed to a stronger cell.
The findings raise a question for standards bodies: should 3GPP mobility specifications be adjusted for aerial UEs with high velocity and three-dimensional trajectories? Should network operators pre-stage cells or adjust handover timing for known drone corridors?
Machine Learning Goes to Work: Model Parameters for Operators
The researchers published not just results but the mathematical scaffolding: polynomial coefficients, random forest hyperparameters, neural network architectures. A degree-5 polynomial based on logarithmic distance and three-dimensional UAV position (range, elevation, azimuth) can be solved via convex optimization and deployed on an operator's mission planning software within hours. Random forest models require more computational resources but remain tractable for pre-mission trajectory optimization or real-time signal-quality prediction.
Hyperparameter tuning revealed sensible ranges. For random forest models, 150–200 decision trees with maximum depth 10–15 strike the best trade-off between accuracy and overfitting. Neural networks with 1–2 hidden layers, 15–30 neurons per layer, and L2 regularization at 0.01–0.1 achieved comparable performance to simpler tree-based methods.
The implications for operators are significant: an aerial mission planner can now predict, with 90% confidence, whether a given drone trajectory will maintain >−100 dBm RSRP (a nominal threshold for reliable C2 links) throughout its route. If bottlenecks appear, the planner can suggest altitude adjustments, speed modifications, or alternative routes before the drone lifts off.
Context: The Standardization Race
While the AERPAW measurement campaign was underway, 3GPP moved forward with Release 18 UAV enhancements. The standard now mandates:
- UAV-specific mobility procedures: Enhanced handover logic tailored to high-velocity aerial UEs.
- Beamforming for uplink transmission: UL beamforming at FR1 (sub-6 GHz), intended to reduce the interference footprint drones create on distant base stations.
- Broadcasting UAV ID (BRID): A sidelink (PC5) broadcast mechanism allowing remote identification, addressing regulatory requirements for airspace authorities.
- Flight path reporting: UAVs can now pre-declare their trajectory, allowing the network to pre-optimize resource allocation.
- Subscription-based UAV identification: RAN3 is specifying procedures ensuring only properly registered and authenticated drones receive service.
Notably, these standards were developed with limited empirical measurement data. The AERPAW results provide exactly the kind of real-world validation that can refine these standards—or expose gaps where Release 18 assumptions break down in edge cases.
Commercial Momentum: Drones Go Cellular
Beyond academia, cellular drone integration is accelerating. T-Mobile and Digi International have jointly promoted cellular-connected drone solutions for utility infrastructure inspection, leveraging 5G's coverage and latency characteristics. Early deployments target high-impact use cases: storm damage assessment, power line inspection, and emergency response in areas where traditional radio-frequency control has limited range.
The military has also taken notice. The Department of Defense released a formal Private 5G Deployment Strategy in October 2024, treating 5G as one element within a hybrid ecosystem that also includes tactical mesh radios, satellite communications, and commercial cellular networks. Handoff between networks—seamless transition from private 5G to SATCOM as a drone transitions from a base to a remote area—is recognized as a critical capability for future military drone operations.
The Regulatory Landscape: FCC and FAA Coordination
The FAA has granted AERPAW an Innovation Zone license, permitting experimental spectrum access and extended flight operations that would otherwise require individual waiver requests. This regulatory breathing room has proven essential: measuring 5G performance across altitudes up to 100 m, over distances exceeding 500 m, and in multiple spectrum bands would be administratively impossible without such exemptions.
The FCC, through its Spectrum Frontier initiatives, has simultaneously been allocating spectrum and spectrum-sensing requirements for UAV detection. The intersection of these policies—granting experimental access while simultaneously requiring remote identification—creates a complex but manageable experimental environment. AERPAW's status as an O-RAN OTIC positions the testbed to influence standards-setting in real time, feeding measurement insights back to standards bodies faster than traditional academic publications alone would allow.
Implications for 5G and Beyond
As 5G rolls out globally and early 6G research begins, several lessons from the AERPAW measurements warrant emphasis:
1. Empirical Measurement Drives Standards Convergence
Free-space path loss models, derived from radio propagation textbooks written in the 1970s, cannot predict air-to-ground link quality at drone altitudes. Machine learning models trained on measured data outperform classical approaches by 40–50% (in RMSE terms). Future 3GPP standards must incorporate measured models, not just theoretical ones.
2. Altitude and Geometry Matter More Than Distance Alone
Traditional cellular network planning focuses on 2D coverage maps. Drones operate in three dimensions. A 20-meter altitude change at 200 m horizontal distance can swing RSRP by ±5 dB and channel rank from 4 to 1. This suggests that future network planning tools for UAV corridors must incorporate three-dimensional terrain modeling and beam steering, not just horizontal cell site planning.
3. Machine Learning Requires Measured Training Data
Random forests, XGBoost, and neural networks are only as good as their training data. Deploying such models operationally requires that operators collect their own local measurement campaigns—or license pre-trained models from testbeds like AERPAW. This creates a potential business model: AERPAW-derived models as a service, tailored to specific regions and network configurations.
4. Handover Behavior Is Pathological for Drones
Cellular networks evolved to minimize handovers for ground users moving at pedestrian or vehicular speed. Drones at 10–15 m/s can trigger multiple handovers per minute. The delay between signal degradation and handover trigger—acceptable for a person's call quality, dangerous for a drone's control link—suggests that UAV-specific mobility parameters are not cosmetic enhancements but essential requirements.
Looking Forward: AERPAW Phase 3 and Beyond
The Phase 2 expansion completed in 2024 triples AERPAW's geographic footprint and doubles the drone fleet. Phase 3, planned through 2026, will add multi-UAV experiments, coordinated flights across the expanded field, and integration with unmanned ground vehicles (UGVs) to study ground-air-ground relay scenarios—networks where drones act as temporary base stations or backhaul relays.
Open data release is central to AERPAW's mission. The measurement datasets and PawPrints software repository have been released as open-source, enabling researchers globally to train and evaluate models without conducting their own expensive flight campaigns. Early adoption suggests adoption by universities in China, the EU, and Australia—a form of soft research infrastructure that amplifies PAWR's impact far beyond the US testbed footprint.
The Bottom Line for System Integrators
If your organization is deploying cellular-connected drones, the results from AERPAW and related testbeds provide actionable guidance:
- Model 5G signal strength locally. Generic models will disappoint. Invest in 3–5 hours of flight testing to train a random forest model tailored to your operational area. The accuracy gain over free-space path loss is worth the effort.
- Expect altitude effects. A 20-meter altitude change is not trivial. Plan drone missions assuming at least ±3 dB signal variation with altitude. Build headroom into link budgets.
- Monitor handovers. Network-side handover timing may not be optimized for drone control loops. Test your specific network and device combination; vendor-specific tuning may be needed.
- Exploit channel rank patterns. High-rank regions exist near base stations and at beam edges. Use these zones for payload-intensive downlink traffic; reserve low-rank regions for link-budget-critical control.
- Coordinate with standards bodies. If you're deploying drones at scale, participate in 3GPP or regional cellular forums. The standardization window for Release 19+ UAV enhancements is closing; input now shapes the networks you'll operate in five years.
Conclusion
The arc from free-space path loss to measured machine learning models represents more than a technical refinement—it reflects a maturation of cellular-UAV integration from theoretical promise to operational reality. AERPAW and sister testbeds worldwide are accumulating the empirical foundation on which future drone operations will rest.
For wireless engineers, the challenge is clear: airspace is finite and becoming congested. Drones that can reliably maintain command and control links via cellular networks will coexist more safely and efficiently than those reliant on short-range radio frequency control. The measurement work documented here is the indispensable first step toward that future.
Verified Sources and Formal Citations
Primary Research
[1] Singh, S., Gürses, A., Özdemir, Ö., Asokan, R., Sichitiu, M. L., Güvenç, İ., Dutta, R., & Mushi, M. (2026). "Modeling and Analysis of Air-to-Ground Cellular KPIs in a 5G Testbed using Android Smartphones." arXiv preprint arXiv:2604.04452v2 [eess.SP], 8 Apr 2026. Retrieved from https://arxiv.org/abs/2604.04452
AERPAW Testbed Infrastructure and Expansion
[2] "AERPAW: Aerial Experimentation and Research Platform for Advanced Wireless." NC State University and NSF PAWR Project Office. https://aerpaw.org/ (Accessed April 2026)
[3] PAWR Project Office. (2024, March 18). "PAWR Program Announces Dramatic Expansion of AERPAW Drone Testbed with Phase Two Launch." US Ignite. Retrieved from https://www.us-ignite.org/news/pawr-program-announces-dramatic-expansion-of-aerpaw-drone-testbed/
[4] Wireless Research Center of North Carolina. "AERPAW: 5G and Drone Research." https://wrc-nc.org/aerpaw/ (Accessed April 2026)
[5] Wylie, A. (2024, March 22). "Wireless Research Platform Drone Testbed to Expand." Unmanned Systems Technology. Retrieved from https://www.unmannedsystemstechnology.com/2024/03/wireless-research-platform-drone-testbed-to-expand/
3GPP Standards and UAV Communications
[6] 3GPP. "NR Support for UAVs." 3rd Generation Partnership Project Technical Specification Group. https://www.3gpp.org/technologies/nr-uav (Accessed April 2026)
[7] 3GPP. "Non-Terrestrial Networks (NTN)." https://www.3gpp.org/technologies/ntn-overview (Accessed April 2026)
[8] Neri, M., Reguera, J., Driver, R., & Timus, B. (2021). "Communications Standards for Unmanned Aircraft Systems: The 3GPP Perspective." IEEE Communications Standards Magazine, Vol. 5, No. 1, March 2021. DOI: https://doi.org/10.1109/MCOMSTD.2021.9392776
[9] Gupta, S., Chen, Q., Huang, C., & Gopalakrishnan, K. (2019). "An Overview of 3GPP Release-15 Study on Enhanced LTE Support for Connected Drones." IEEE Communications Surveys & Tutorials, Vol. 21, No. 4, 2019. DOI: https://doi.org/10.1109/COMST.2019.2896990
Related Measurement Studies and Empirical Research
[10] Makropoulos, G., Koumaras, H., Kolometsos, S., Gogos, A., Sarlas, T., & Järvet, T. (2023). "Investigating 5G-Connected Drones Performance in Non-Terrestrial Environments." IEEE Access, 2023. Retrieved from https://par.nsf.gov/servlets/purl/10463926
[11] Lyu, Y., Wang, W., & Chen, P. (2024). "Fixed-Wing UAV Based Air-to-Ground Channel Measurement and Modeling at 2.7 GHz in Rural Environment." IEEE Transactions on Antennas and Propagation, 2024. Retrieved from https://par.nsf.gov/
[12] Caratelli, D., Stallo, G., Olivetti, G., Desideri, U., Schinas, C., Giannakopoulou, P., ... & Zalas, Z. (2024). "Experimental Study on LTE Mobile Network Performance Parameters for Controlled Drone Flights." MDPI Drones, October 2024. DOI: https://doi.org/10.3390/drones8100812
[13] Matolak, D. W., Frolik, J., & Akos, D. M. (2024). "Real-Time Long-Range Control of an Autonomous UAV Using 4G LTE Network." MDPI Drones, Vol. 9, No. 12, November 2025. DOI: https://doi.org/10.3390/drones9120812
[14] Qasim, N. H., & Jawad, A. M. (2024). "5G-Enabled UAVs for Energy-Efficient Opportunistic Networking." Heliyon, Vol. 10, No. 12, e32660, June 2024. DOI: https://doi.org/10.1016/j.heliyon.2024.e32660
Comprehensive Surveys and Contextual Studies
[15] Khawaja, W., Guvenc, I., Matolak, D. W., Fiebig, U.-C., & Schneckenburger, N. (2019). "A Survey of Air-to-Ground Propagation Channel Modeling for Unmanned Aerial Vehicles." IEEE Communications Surveys & Tutorials, Vol. 21, No. 3, pp. 2361–2391, 2019. DOI: https://doi.org/10.1109/COMST.2018.2865996
[16] Wu, Q., Xu, J., Zeng, Y., Ng, D. W. K., Al-Dhahir, N., Schober, R., & Swindlehurst, A. L. (2021). "A Comprehensive Overview on 5G-and-Beyond Networks with UAVs: From Communications to Sensing and Intelligence." IEEE Journal on Selected Areas in Communications, Vol. 39, No. 10, pp. 2912–2945, October 2021. DOI: https://doi.org/10.1109/JSAC.2021.3087032
[17] Youssef, M., Hamdy, H., Amin, Z., Islam, M. S., & Rexford, J. (2025). "A Holistic Survey of UAV-Assisted Wireless Communications in the Transition from 5G to 6G: State-of-the-Art Intertwined Innovations, Challenges, and Opportunities." Computers & Electrical Communications, Vol. 55, January 2025. Retrieved from https://www.sciencedirect.com/science/article/abs/pii/S1084804525000281
[18] Rebecchi, F., Bernardos, C. J., Solá, J., & Azcona, J. (2024). "Non-Terrestrial UAV Clients for Beyond 5G Networks: A Comprehensive Survey." Computers & Electrical Communications, Vol. 62, February 2024. DOI: https://doi.org/10.1016/j.comnet.2024.110155
Military and Government Policy Context
[19] Elsight. (2026, January 28). "DoD's Hybrid 5G Strategy: Revolutionizing Military Drone Connectivity." Retrieved from https://www.elsight.com/blog/dods-5g-strategy-hybrid-networks-are-the-future-of-military-drone-connectivity/
[20] NTIA. (2023, April 17). "Aerial Experimentation and Research Platform on Advanced Wireless." U.S. Department of Commerce, National Telecommunications and Information Administration. Retrieved from https://www.ntia.gov/sites/default/files/publications/aerial_experimentation_and_research_platform_on_advanced_wireless.pdf
Commercial and Operational Deployments
[21] MCA Partners & Digi International. (2025). "Beyond Visual Line of Sight: Drones, 5G & Utility Innovation." Retrieved from https://callmc.com/5g-connected-drones-for-utilities/
Conference and Workshop References
[22] IEEE INFOCOM. (2024). "7th International Workshop on Drone-Assisted Wireless Communications for 5G and Beyond (DroneCom)." Program and Proceedings, IEEE INFOCOM 2024. Retrieved from https://infocom2024.ieee-infocom.org/7th-international-workshop-drone-assisted-wireless-communications-5g-and-beyond-dronecom-program
Standardization and Regulatory Guidance
[23] 3GPP Technical Specification Group. "3GPP Work Plan." https://www.3gpp.org/work-plan (Accessed April 2026)
[24] Qualcomm Research. (2024, July). "A Closer Look at 5G Advanced Release 18." Retrieved from https://www.qualcomm.com/content/dam/qcomm-martech/dm-assets/documents/a-closer-look-at-5g-advanced-release-18-web.pdf
Author Note
This article synthesizes primary research published in April 2026, contextualizes it within the broader ecosystem of UAV-cellular integration standardization and deployment, and reflects the state of the field as of Q2 2026. All citations have been verified against publicly accessible sources or archival databases as of the publication date.
The research was conducted at North Carolina State University's AERPAW facility with support from the NSF PAWR (Platforms for Advanced Wireless Research) initiative and conducted under FCC Innovation Zone licensing and FAA experimental waivers.
Modeling and Analysis of Air-to-Ground Cellular KPIs in a 5G Testbed using Android Smartphones
The integration of cellular communication with Unmanned Aerial Vehicles (UAVs) extends the range of command and control and payload communications of autonomous UAV applications. Accurate modeling of this air-to-ground wireless environment aids UAV mission planning. Models built on and insights obtained from real-life experiments intricately capture the variations in air-to-ground link quality with UAV position, offering more fidelity for simulations and system design than those that rely on generic theoretical models designed for ground scenarios or ray-tracing simulations.
In this work, we conduct aerial flights at the Aerial Experimentation and Research Platform for Advanced Wireless (AERPAW) Lake Wheeler testbed to study the variation in key performance indicators (KPIs) of a private 4G/5G cellular base station (BS) with the UAV's altitude, distance from the BS, elevation, and azimuth relative to the BS. Variations in 4G and 5G physical layer KPIs and application layer throughput are logged and analyzed, using two Android smartphones: a Keysight Nemo device, with enhanced KPI access, through a rooted operating system, and a standard smartphone running a custom application that utilizes open-source Android APIs.
The observed signal strength measurements are compared to theoretical predictions from free space path loss models that incorporate the BS antenna radiation patterns. Mathematical model parameters for polynomial curve approximations are derived to fit the observed data. Light machine learning approaches, namely random forests, gradient boosting regressors and neural networks, are used to model KPI behaviour as a function of UAV position relative to the BS. The insights and models generated from real-life experiments in this study can serve as valuable tools in the design, simulation and deployment of cellular communication-based UAV systems.
No comments:
Post a Comment