Low-Altitude Economy Networks (LAENets)

• LAENets are integrated infrastructures combining communication, sensing, computation, and control to support low-altitude UAV and eVTOL operations.
• They employ layered architectural designs, including terminal setups, digital airspace management, and hierarchical coordination across terrestrial, aerial, and non-terrestrial nodes.
• Recent studies show enhanced interference mitigation, energy efficiency, and secure operations using innovations like rotatable antennas, digital twins, and LLM-enhanced optimization.

Low-Altitude Economy Networks (LAENets), also framed in some papers as Low-Altitude Economic Networking, denote the integrated air-ground infrastructure that supports low-altitude operations by UAVs, eVTOLs, and related aerial platforms. Across the recent literature, LAENets are described as multi-functional systems that combine communication, sensing, computation, control, and airspace management for services such as logistics delivery, infrastructure inspection, public safety, environmental monitoring, temporary coverage extension, and urban air mobility. The altitude scope is not uniform across papers: several works center on airspace below 1,000 m, while others describe broader low-altitude regimes below approximately 3,000 m or emphasize specific deployment bands such as below roughly 300 m (Cai et al., 27 May 2025, Wu et al., 15 Sep 2025, Wang et al., 30 Apr 2025).

1. Definition and conceptual scope

Recent work treats LAENets not as a single protocol stack but as an infrastructural category. One formulation defines them as “a ubiquitous, intelligent, and interoperable network infrastructure capable of seamlessly connecting unmanned aerial vehicles (UAVs) with terrestrial digital ecosystems, as well as providing airspace and security management.” Another frames them as the integrated networking fabric that combines 6G non-terrestrial networks and terrestrial networks into an integrated air-ground network, with distinct control and non-payload communication and payload communication functions. A further formulation treats them as the multi-layered operational fabric that integrates communication, computing, sensing, navigation integrity, and regulation, with High-Altitude Platforms (HAPs) acting as pivotal regional nodes (Luo et al., 10 Jun 2025, Jiang et al., 2023, Huang et al., 23 Feb 2026).

This literature consistently distinguishes LAENets from earlier UAV-centric communication systems. Traditional aerial systems are commonly described as emphasizing air-to-ground communication services, whereas LAENets require integrated sensing, computation, control, energy-delivering functions, and traffic management. They are also distinguished by dense flying nodes, stronger coupling to terrestrial infrastructures, increased airspace management demands, and heightened vulnerability to cyberattacks and interference in open-air environments (Wu et al., 15 Sep 2025, Wang et al., 30 Apr 2025).

The channel and interference environment is correspondingly treated as intrinsically three-dimensional. Elevation-dependent line-of-sight models recur across the literature, for example

$$P_{\mathrm{LoS}}(\theta)=\frac{1}{1+a\exp\!\left(-b(\theta-a)\right)},$$

with path-loss and shadowing models used to capture the joint growth of coverage and interference under LoS-dominant aerial links. This modeling choice reflects a central conceptual point: LAENets are not merely wider-area cellular systems, but altitude-sensitive networks in which mobility, blockage, beam geometry, and surveillance obligations remain tightly coupled (Jiang et al., 2023).

2. Architectural strata and network elements

A recurring architectural pattern is layering. One core architecture divides LAENets into an Airborne Terminal and Physical Infrastructure Layer, an Intelligent Collaboration and Digital Airspace Layer, and a Multi-Collaboration and Service Assurance Layer. In that formulation, airborne terminals integrate power, sensing, navigation, and flight-control subsystems, while the physical layer includes takeoff and landing stations, transfer facilities, energy stations, emergency landing sites, secure parking, and maintenance support. The upper layers add spatiotemporal resource management, City Information Modelling, knowledge and rule bases, supervision, enterprise-oriented flight services, and flight-control functions such as flow management and real-time traffic control (Wang et al., 30 Apr 2025).

A second architectural strand is cellular-native and explicitly control-data decoupled. UTICN organizes LAENets into five functional subsystems—sensing, positioning, communication, management, and service—and separates the UAV management network from the public internet. Its management plane includes UAGF, UACF, UARF, UAMF, and UANF; edge UAGF, MEC, UPF, and network slicing support full lifecycle management from registration to inter-domain handover, plus emergency override actions such as forced hovering, return-to-home, and landing (Luo et al., 10 Jun 2025).

A third strand is hierarchical across altitude and administrative scope. In the HAP-centered view, the global tier consists of LEO satellites and cloud systems, the regional tier consists of HAPs at around 20 km altitude, and the local tier consists of ground nodes and UAV meshes. HAPs are described as regional “super nodes” providing low-latency access, MEC, integrated sensing and communication, navigation integrity support, and supervisory regulation; a single HAP can sustain “private-network-grade” support across 200–500 km regions. Related NTN-oriented work similarly organizes LAENets around non-terrestrial nodes, terrestrial base stations, digital management systems, CNPC, payload communication, and sensing planes (Huang et al., 23 Feb 2026, Jiang et al., 2023).

Taken together, these designs suggest an architectural convergence around layered service assurance, separated control and mission data, and hierarchical coordination across terrestrial, aerial, and extra-aerial domains. This suggests that “network” in LAENets is best read as a composite of radio access, sensing, computing, and governance rather than as a purely communications-layer construct.

3. Integrated technical functions

Communication remains foundational, but recent work treats it as inseparable from beam control, sensing, and resource coordination. A representative example is the rotatable antenna system (RAS), introduced to address the limits of terrestrial base stations with down-tilted, fixed-sector antennas. RAS adds mechanical or electronic boresight control, enabling flexible beamforming in three dimensions, extended low-altitude coverage, and improved interference mitigation. In prototype experiments, mechanically driven RAS yielded a clearer, more regular constellation and approximately $7$ dBm higher received signal power than a fixed antenna; simulations further reported higher signal-to-clutter-and-noise ratio and more uniform received power across azimuth (Li et al., 1 Nov 2025).

Spectrum and radio-environment awareness are likewise elevated to core functions. A two-stage generative-AI framework for temporal spectrum cartography uses a Reconstructive Masked Autoencoder to recover full radio-frequency power maps from sparse, temporally varying measurements, then uses a Multi-agent Diffusion Policy to steer UAV sensors toward informative regions. In the reported experiments, reconstruction error was reduced by $57.35\%$ relative to Kriging and $88.68\%$ relative to an autoencoder baseline; the trajectory planner further improved cumulative mean-squared error and stability in dynamically evolving low-altitude environments (Zhao et al., 21 May 2025).

Computation is no longer treated merely as backend support. One line of work models UAV-enabled LAENets as platforms for onboard vision-language inference, where uplink transmission, image resolution, answer generation speed, and UAV trajectory are jointly optimized. In that setting, a hierarchical framework combining Alternating Resolution and Power Optimization with an LLM-augmented reinforcement-learning trajectory planner reduces worst-case latency by about $45\%$ relative to random resolution, power, and trajectory selection, by about $25\%$ relative to a geometric heuristic, and by an additional $13.7\%$ relative to a manual-reward PPO baseline (Li et al., 11 Oct 2025).

A more market-oriented extension treats airborne compute as a tradable resource. In Low-Altitude Computility Networks, the onboard computing utility of drones and eVTOLs is tokenized as Real-World Assets and coordinated through a hybrid blockchain fabric. In the urban logistics case study, a fungible COMP token denotes a unit of computation, conceptually $1$ GFLOP; permissioned chains handle identity and confidential operations, while permissionless chains host marketplace logic, token issuance and settlement. This shifts LAENets toward a model in which communication, trust, and edge computing are jointly orchestrated rather than separately provisioned (Luo et al., 25 Aug 2025).

4. Optimization, learning, and digital-twin control

Optimization in LAENets is commonly formulated as sequential decision-making under uncertainty. A standard formulation models the UAV-environment interaction as an MDP with policy objective

$$J(\pi)=\mathbb{E}\!\left[\sum_{t=0}^{\infty}\gamma^t r_t\right].$$

Within this framework, LLM-enhanced reinforcement learning has been organized into four roles: information processor, reward designer, decision-maker, and generator. In a UAV-assisted IoT case study, the use of an LLM-designed reward function reduced energy consumption relative to a manually designed reward, with TD3 achieving up to $7.2\%$ lower final energy and up to $7$0 lower energy at $7$1 Mbits (Cai et al., 27 May 2025).

Optimization pipelines increasingly combine model formulation and policy learning. In low-carbon multi-UAV MEC, a HybridRAG-based LLM agent combines KeywordRAG, VectorRAG, and GraphRAG to formulate a carbon-minimization problem over offloading, computing, and trajectory variables, then solves it using a Double Regularization Diffusion-enhanced Soft Actor-Critic algorithm. The reported results show that the resulting policy exhibits a $7$2 performance improvement over SAC, while CodeCarbon measurements put training emissions at about $7$3 g and inference at about $7$4 g per decision (Wen et al., 19 Jun 2025).

Digital twins provide another optimization stratum, especially where communication objectives are tightly coupled. DT-MOO models candidate configurations through a synchronized spectrum twin, traffic twin, and interference twin, and evaluates them under joint RSRP–SINR criteria. Its coverage objective is explicitly written as

$7$5

In field validation on a 5G-enabled low-altitude communication network, DT-MOO increased the high-quality coverage rate from $7$6 to $7$7 across all evaluated altitudes relative to an operator-provisioned baseline (Huang et al., 20 Apr 2026).

Competitive multi-provider settings add a game-theoretic layer. In a service-centric deployment model with multiple service providers, an authenticity-guaranteed sealed-bid auction is coupled to resilient federated reinforcement learning. The resulting DAPCR-FedPG method consistently identified $7$8 good nodes out of $7$9 in a scenario with $57.35\%$0 Byzantine nodes, achieved training loss near zero, and stabilized total rewards around $57.35\%$1, supporting the claim that LAENet optimization increasingly spans mechanism design, federated coordination, and adversarial robustness rather than isolated PHY/MAC tuning (Yang et al., 5 Mar 2026).

5. Security, privacy, and covert operation

The physical-layer security literature treats LAENets through the confidentiality-availability-integrity triad. Standard metrics include secrecy capacity, secure outage probability, SINR under jamming, and energy efficiency; one representative secrecy expression is

$57.35\%$2

Surveyed countermeasures include joint trajectory-power optimization, cooperative jamming, RIS-assisted beamforming, physical-layer authentication, anti-jamming and anti-spoofing techniques, GNSS spoofing mitigation, anomaly detection, and injection suppression. A recurrent premise is that LoS-dominant aerial channels improve both legitimate connectivity and adversarial observability, so security cannot be treated as an afterthought (Cai et al., 12 Apr 2025).

Generative-AI methods have been proposed specifically for robustness under uncertainty. A diffusion-based reinforcement-learning framework with a Mixture-of-Experts transformer actor has been used for robust secure beamforming in the presence of channel uncertainty, AoA errors, and covert eavesdroppers. In the reported beamforming case study, the method achieved more than a $57.35\%$3 increase in worst-case achievable secrecy rate relative to benchmark methods, while chance-constrained formulations were described as more balanced than strictly worst-case robust designs (Zhao et al., 25 Feb 2025).

Covertness has also been integrated with ISAC and MEC. In a networked ISAC system toward LAE, an MEC server coordinates multiple APs that both receive UAV computation tasks and illuminate a sensing target, while AP-transmitted dual-functional waveforms also act as friendly jamming against wardens. The work derives closed-form expressions for the detection error probability at wardens and minimizes total energy consumption by jointly optimizing communication, sensing, computation resources, and UAV trajectories through alternating optimization, SCA, and trust-region methods. A related semantic-communication line regulates semantic entropy under low-probability-of-detection constraints through abstraction-level selection, contract theory, and a Regularized Diffusion-based Soft Actor-Critic algorithm; in that setting, RDSAC improved average reward by $57.35\%$4 over SAC and $57.35\%$5 over PPO (Mao et al., 24 Jul 2025, Liu et al., 2 Mar 2026).

Privacy-preserving perception introduces an additional security dimension. In vision-aided ISAC, masked De-Diffusion models extract semantic tokens containing agent type, activity class, and heading orientation while suppressing sensitive visual content. Synthetic reconstructions parsed by YOLOv11 and SlowFast then drive RAT assignment, beamforming, and power control; the resulting DeDiff-VARARO pipeline achieves performance within $57.35\%$6 of a raw-image upper bound while preserving user privacy and scalability in dense environments (Gao et al., 2 Jul 2025).

6. Airspace governance, standardization, and emerging trajectories

Operationalization depends on airspace management and standards as much as on algorithms. UTICN exemplifies this by tying sensing, positioning, communication, management, and service subsystems to city-level supervision and multi-frequency collaborative ISAC. Its reported field metrics include $57.35\%$7 ms air-interface latency, $57.35\%$8 Gbps throughput, $57.35\%$9 m range resolution, successful detection of UAVs with radar cross-section $88.68\%$0, A2X latency below $88.68\%$1 ms, and range greater than $88.68\%$2 km in open environments (Luo et al., 10 Jun 2025).

Standards-aware architectural work places LAENets inside a concrete interoperability agenda. IEEE 1939.1-2021 specifies a scalable low-altitude airspace framework with $88.68\%$3 grid-based route planning; IEEE 1937.8-2024 requires base-station switching delay $88.68\%$4 ms for seamless communication terminal integration; IEEE 1937.3-2024 defines flight-monitoring data with sampling intervals $88.68\%$5 seconds; IEEE P1954, together with IEEE P1920.1 and the IEEE 1900 series, targets self-organizing, spectrum-flexible UAV-swarm communication (Wang et al., 30 Apr 2025).

Airspace structuring is therefore treated as a design variable, not merely a regulatory boundary. The LAWN framework surveys pipeline airspace, corridor airspace, layered altitude strata, blocky three-dimensional cells, and free airspace, alongside capacity analysis through Macroscopic Fundamental Diagrams. This broadens LAENets from a connectivity problem into an airspace-service allocation problem in which routing, separation, capacity, and service classes are co-designed (Wu et al., 15 Sep 2025).

Deployment research increasingly emphasizes testbeds that close the simulation-to-reality gap. AERPAW combines digital-twin, sandbox, and outdoor testbed modes for AI-driven LAE networking. In its federated spectrum-sensing studies, FedSNR delivered up to a $88.68\%$6 accuracy improvement over centralized CNN/LSTM baselines, outperformed FedAvg in low-SNR and low-client regimes, and saturated near $88.68\%$7 accuracy with $88.68\%$8 UAVs. These results are presented not only as algorithmic gains but as evidence that realistic measurement platforms are necessary for iterative refinement and standardization (Tekbıyık et al., 2 Jun 2025).

Future directions are correspondingly multi-layered. HAP-centered work proposes a five-stage roadmap from aerial infrastructure bases to edge–air–cloud closed-loop autonomy; survey work points to digital twins, lightweight edge LLMs, and standardization; architecture-oriented studies add intelligent and adaptive optimization, sustainable energy and power management, generative governance, and three-dimensional airspace coverage. This suggests that LAENets are evolving toward globally connected, regionally orchestrated, and locally adaptive infrastructures whose defining challenge is no longer radio access alone, but the coordination of communication, sensing, computing, control, and regulation at scale (Huang et al., 23 Feb 2026, Wu et al., 15 Sep 2025, Wang et al., 30 Apr 2025).