5G-Enabled UAV Control Applications: Experimental Evaluation of Network and Computing Architectures

Despite the increasing availability of 5G networks, there is still limited experimental evidence comparing public and private deployments, mmWave communications, and distributed processing architectures for real UAV applications.

This paper presents an experimental evaluation of 5G-enabled UAV communications through a campaign comprising 27 flights comparing public and private 5G deployments, including sub-6 GHz and mmWave operation in the latter. In addition to network characterization experiments (12 flights), a vision-based UAV application (15 flights) is evaluated under three computing architectures: onboard processing, Multi-access Edge Computing (MEC), and remote cloud computing. The study analyzes network-level measurements, including throughput, latency, and packet loss, flight-related geometric factors, such as altitude and distance to the serving antenna, and application-level metrics, such as image-processing time and end-to-end latency. The influence of flight-related geometric factors on network performance is statistically assessed using Spearman's rank correlation to quantify associations between geometric and network variables, complemented by permutation-based feature importance to estimate the relative contribution of each factor to the observed network performance.

The experimental results show that private 5G deployments provide significantly more stable communication conditions than public networks, enabling lower latency and higher throughput for latency-sensitive UAV applications. Furthermore, mmWave communications substantially improve downlink throughput and latency under favorable propagation conditions, while their benefits in the uplink remain limited due to practical deployment constraints. Regarding computing architectures, the results highlight the trade-off between communication dependency and computational capacity. Although cloud-based processing introduces additional communication delay, its substantially lower image-processing time compensates for this overhead, resulting in the lowest overall end-to-end latency among the evaluated distributed architectures. Nevertheless, its performance remains strongly dependent on network conditions.

By jointly evaluating public and private 5G deployments, sub-6 GHz and mmWave communications, flight-related geospatial factors, and different computing architectures under real flight conditions, this work provides a comprehensive experimental characterization for 5G-enabled UAV applications.