LACNets: Integrated Low-Altitude Computility Networks

• LACNets are integrated low-altitude networks that jointly virtualize communication, computation, sensing, and control as a unified service fabric.
• They fuse aerial, terrestrial, and cloud–edge resources to enable applications like urban air mobility, emergency response, and logistics with dynamic scheduling.
• Advanced techniques such as ML-driven spectrum sensing, RL-based resource allocation, and digital twin optimization underpin their robust performance under uncertainty.

Low-Altitude Computility Networks (LACNets) are low-altitude airspace infrastructures in which communication, computation, sensing, control, and related utility functions are jointly virtualized, orchestrated, and optimized across aerial platforms, terrestrial infrastructure, and cloud–edge resources. In published formulations, they appear as cloud–edge–air integrated networks and as the natural evolution of AI-driven low-altitude economy networks, operating in low-altitude airspace typically below $1000$ m in some formulations and below $3000$ m in others, and serving urban air mobility, emergency response, logistics, environmental sensing, surveillance, and allied services (Sun et al., 28 Sep 2025, Tekbıyık et al., 2 Jun 2025). Their distinguishing feature is that computing power is treated as a routable, schedulable, and, in some formulations, tradable resource rather than as a peripheral endpoint capability, so that UAVs, eVTOLs, balloons, and related platforms act simultaneously as communication nodes, sensors, control agents, and edge computing resources (Wu et al., 15 Sep 2025, Luo et al., 25 Aug 2025).

1. Conceptual lineage and defining characteristics

LACNets emerge at the intersection of several research lines. Low-Altitude Economy networks frame the airspace as an operational domain for logistics, emergency rescue, air taxis, precision agriculture, environmental sensing, surveillance, and monitoring. Low-Altitude Wireless Networks extend this by integrating communication, sensing, computation, control, wireless power transfer, and low-altitude airspace structuring and air traffic management into a unified design. Computing Power Networks contribute a service-centric view in which distributed computing, networking, and storage resources are jointly abstracted and scheduled. LACNets synthesize these strands by making computation a first-class network utility in low-altitude airspace rather than an auxiliary function at endpoints (Wu et al., 15 Sep 2025, Sun et al., 28 Sep 2025).

A recurring definition in the literature characterizes LACNets as cloud–edge–air integrated networks in which low-altitude communication, computing, and storage resources are jointly virtualized, orchestrated, and optimized as a single computing-power-aware service fabric for low-altitude applications. Related formulations describe them as AI-orchestrated, multi-layer low-altitude networks in which aerial platforms, terrestrial infrastructure, and cloud or edge resources jointly deliver communication, computation, and utility functions under constraints on spectrum, energy, latency, and safety (Sun et al., 28 Sep 2025, Tekbıyık et al., 2 Jun 2025). The term computility is explicitly used to denote the computing utility of aerial devices, namely usable computational capacity such as CPU or GPU cycles, memory, and sensor-processing capability that can be offered as a service (Luo et al., 25 Aug 2025).

Several misconceptions are directly addressed by this conceptual lineage. LACNets are not merely UAV communication networks, because traditional aerial systems predominantly focus on air-ground communication services while neglecting the integration of sensing, computation, control, and energy-delivering functions. They are also not simply classic edge or fog systems extended upward, because classic MEC or fog deployments are mostly static, whereas LACNets place aerial platforms into the scheduling loop as mobile compute, sensing, and control resources with dynamic topology, frequent handovers, and trajectory-dependent connectivity (Wu et al., 15 Sep 2025, Sun et al., 28 Sep 2025).

2. Architecture, planes, and service organization

The architectural vocabulary of LACNets is notably layered. One line of work proposes a three-layer low-altitude economy architecture consisting of an airborne terminal and physical infrastructure layer, an intelligent collaboration and digital airspace layer, and a multi-collaboration and service assurance layer. Another describes a cloud–edge–aerial integration paradigm with an aerial layer, edge layer, regional or central cloud layer, and a programmable control and data plane spanning ground and air. A related low-altitude wireless network formulation organizes the system as tightly coupled data, control, sensing, and computing planes (Wang et al., 30 Apr 2025, Sun et al., 28 Sep 2025, Yuan et al., 14 Jun 2025).

Stratum	Representative elements	Main functions
Physical air–ground infrastructure	UAVs, eVTOLs, balloons, HAPs, BSs, takeoff and landing stations, energy stations	connectivity, sensing, mobility, actuation
Functional planes	data plane, control plane, sensing plane, computing plane, CPN control plane	routing, C2, offloading, in-network computing, policy execution
Digital airspace and services	CIM, knowledge base or rule base, supervision, flight service, flight control	spatiotemporal resource management, compliance, scheduling, traffic control

Within this architecture, the aerial layer contains heterogeneous UAVs, eVTOL air taxis, drones, balloons, and, in some formulations, HAPs. These act as sensing nodes, communication nodes such as aerial base stations or relays, and computing nodes with onboard processors or DPUs/FPGAs for in-network or task-oriented computing. The edge layer contains BS-attached edge servers, micro-data centers, roadside units, and UGV-mounted MEC; the regional and central cloud layer provides large-scale data centers for global orchestration, training, transfer learning, and digital-twin simulation (Sun et al., 28 Sep 2025).

The logical planes are equally important. The data plane carries user traffic, sensing data, and compute tasks. The control plane carries flight control messages, telemetry, spectrum policies, AI model updates, and policy synchronization. The sensing plane supplies environmental awareness through cameras, LiDAR, mmWave radar, IMU, and RF sensing. The computing plane distributes processing across onboard processors, edge servers, and cloud infrastructure for obstacle detection, scene understanding, prediction, task offloading, and long-term analytics (Yuan et al., 14 Jun 2025).

Digital airspace components make this architecture operational. City Information Modelling provides unified 3D urban models for path planning and obstacle avoidance. Knowledge bases and rule bases encode meteorology, operating rules, and regulations. Low-altitude supervision systems handle identity authentication, certificates, biometrics, aircraft codes, and violation handling. Flight service systems optimize routes, timing, and resource coordination, while flight control systems perform flow management and issue real-time commands such as takeoff, landing, and altitude or course changes (Wang et al., 30 Apr 2025). A plausible implication is that LACNets should be understood not only as communication infrastructures but also as digital airspace systems.

3. Core technical mechanisms: spectrum intelligence, ISCC, and multi-tier offloading

A central LACNet mechanism is machine-learning-based spectrum sensing and coexistence. One concrete implementation uses federated learning for UAV-based spectrum sensing: each UAV runs a lightweight CNN on received complex IQ samples, with real and imaginary parts processed by 1D convolutions, batch normalization, dropout, and global average pooling. Local models are trained on-device, and raw IQ data are not shared. The global aggregation rule, termed FedSNR, weights model updates by node SNR rather than using simple FedAvg: $$w_{\text{new}} = \sum_{k=1}^{K} \alpha_k w^{(k)}, \qquad \alpha_k = \frac{\mathrm{SNR}_k}{\sum_{j=1}^{K}\mathrm{SNR}_j}.$$ Reported outcomes include up to $\approx 15\%$ accuracy improvement over centralized CNN/LSTM baselines, accuracy stabilizing around $97.6\%$ with 16 UAVs, and clear gains over FedAvg, particularly in low-SNR regimes (Tekbıyık et al., 2 Jun 2025).

The same research line places RL and DRL at the center of resource allocation and trajectory control. In the stated MDP formulation, the state may include UAV position, velocity, altitude, CSI, user locations, traffic demand, and remaining energy; actions may include trajectory adjustments, transmit power $P_t$, band selection, and task offloading. A typical reward takes the form

$$r_t = \lambda_1 R_t - \lambda_2 L_t - \lambda_3 E_t - \lambda_4 I_t,$$

balancing throughput, latency, energy consumption, and interference. The literature further associates this control layer with digital twin networks, GRUs plus Value Decomposition Networks, apprenticeship learning via inverse RL, and federated model updates under non-IID mobility conditions (Tekbıyık et al., 2 Jun 2025).

Integrated sensing, communication, and computing is another core mechanism. A full-duplex UAV-enabled low-altitude platform has been studied as a single-node realization of this principle: the aerial node acts simultaneously as a wireless access point or relay, a radar, and a mobile edge server, and jointly optimizes task allocation, computation resource allocation, transmit beamforming, and receive beamforming. In the cited numerical study, the proposed alternating-optimization scheme saves up to $54.12\%$ energy consumption performance compared to benchmark schemes (Chen et al., 25 Apr 2025). This makes explicit that a LACNet node can be treated as a joint ISCC optimizer rather than as a separated radio or MEC element.

At larger scales, satellite-assisted designs extend the same logic. A distributed satellite MIMO architecture formed by multiple visible LEO satellites and a distributed LAV MIMO architecture formed by LAV fleets are combined with a two-timescale optimization scheme. In the reported case study, distributed LAV MIMO reduces the required LAV transmission power by $>85\%$ compared to a single LAV for a target rate of $18$ bps/Hz, while distributed satellite MIMO extends service duration up to $3000$0 that of co-located satellite MIMO and yields fewer handovers (He et al., 7 May 2025). This places LACNets within a broader space–air–ground computility continuum.

4. Formal models, predictive control, and robust optimization

The formal apparatus used for LACNets is dominated by latency decomposition, AoI, queueing, graph models, MDPs, and robust optimization. A standard end-to-end latency decomposition in cloud–edge–air architectures is

$3000$1

where transmission, queueing, computation, and feedback delays all contribute to service completion time. This decomposition supports computation-aware routing, offloading, and service function chain placement across aerial, edge, and cloud nodes (Sun et al., 28 Sep 2025).

Age of Information is a recurring metric because many low-altitude services are state-update or control-loop dominated rather than bulk-throughput dominated. In representative formulations,

$3000$2

where $3000$3 is the timestamp of the latest update for target $3000$4. AoI appears in aerial sensing, inspection, emergency, and security-aware ISCC optimization, and it is explicitly coupled to computing rates and secrecy-aware transmission rates in GI/M/1-type or M/M/1-type models (Sun et al., 28 Sep 2025, Wang et al., 3 Nov 2025).

Robustness to uncertainty is addressed through distributionally robust optimization. In one low-altitude network architecture with UAVs and a HAP jointly providing computation offloading, uncertain task sizes are modeled by ambiguity sets

$3000$5

and the system minimizes worst-case delay by jointly optimizing offloading decisions and UAV trajectories. The solution combines an outer-layer convex problem over probability distributions with an inner-layer mixed-integer problem solved via Benders decomposition, while UAV trajectories in the subproblem are optimized by successive convex approximation (Jiang et al., 27 Oct 2025). This suggests that LACNets require explicit uncertainty modeling rather than deterministic planning whenever task demand, mobility, or channel conditions are imperfectly known.

Predictive communications provide a complementary control doctrine. The predictive framework fuses pre-filed mission trajectories with stable large-scale radio maps and decomposes resource management into strategic routing, tactical timing, and operational power control. The reported benefit is an order-of-magnitude reduction in cross-tier interference (Chen et al., 1 Sep 2025). A plausible implication is that LACNets should align prediction fidelity with decision timescale: coarse, long-range prediction for route and path reservation; finer local prediction for timing and re-routing; and short-range adaptation for power control.

5. Experimental platforms, digital twins, and representative deployments

Testbed-driven validation has become a defining methodological feature of LACNet research. AERPAW is described as a platform with programmable UAVs and ground SDR nodes deployed across mixed urban–rural environments including Lake Wheeler and the NC State campus. It supports three experimental modes—digital twin, sandbox, and physical testbed—and is used to evaluate FL-based spectrum sensing, DRL-based trajectory and resource allocation, and integrated communication–computing–sensing workflows. Measured altitude-dependent spectrum occupancy shows significant improvement at higher altitudes, for example in Band 13 above 80 m, and heatmaps reveal altitude- and environment-dependent occupancy in LTE, ISM, and n41 bands (Tekbıyık et al., 2 Jun 2025).

Digital twins are also used for deployment optimization. In a 5G-enabled low-altitude communication network, the DT-MOO framework constructs a high-fidelity virtual replica integrating realistic environmental models, electromagnetic propagation, and traffic dynamics, and uses it for coverage–interference co-optimization. In the reported real-world validation, DT-MOO increases the high-quality coverage rate from $3000$6 to $3000$7 across all evaluated altitudes compared to an operator-provisioned, experience-based baseline. Under a stricter SINR threshold $3000$8 dB, the fraction of voxels satisfying the threshold improves from $3000$9 to $$w_{\text{new}} = \sum_{k=1}^{K} \alpha_k w^{(k)}, \qquad \alpha_k = \frac{\mathrm{SNR}_k}{\sum_{j=1}^{K}\mathrm{SNR}_j}.$$0, despite local spatial trade-offs (Huang et al., 20 Apr 2026). These results indicate that cross-objective interactions in 3D low-altitude deployment are sufficiently strong that repeated field tuning is inefficient relative to digital-twin-guided optimization.

The application space is correspondingly broad. Autonomous aerial transportation assigns real-time trajectory control to edge nodes and keeps safety-critical perception onboard. Aerial sensing and inspection use UAV swarms for high-resolution imaging and semantic compression, with edge nodes performing feature extraction and anomaly detection and cloud systems handling historical analytics. Emergency response uses UAVs as ad hoc communication nodes and rapid-deployment sensors, while logistics and parcel delivery combine route planning, authentication, billing, and fleet management into orchestrated service chains (Sun et al., 28 Sep 2025). A plausible implication is that LACNets should be viewed less as a single service network than as a shared substrate for multiple mission classes with heterogeneous QoS, reliability, and timeliness requirements.

6. Security, governance, economic orchestration, and research directions

Security is a primary concern because low-altitude operation brings physical openness, predictable trajectories, frequent mobility, and heavy use of unlicensed spectrum. The threat model includes detection attacks, eavesdropping, jamming, and tampering. Large AI model-enabled secure communications address this by using LLMs and related foundation models to generate semantically enriched state features and intrinsic rewards on top of handcrafted representations, thereby improving RL-based secure communication control. In the reported case study, LLM-enhanced SAC and LLM-enhanced TD3 achieve higher cumulative rewards, faster convergence, and better stability than vanilla SAC, DDPG, and TD3 in a scenario involving an aerial autonomous vehicle, an aerial eavesdropper, and a ground jammer (Zhang et al., 1 Aug 2025).

Security-aware ISAC optimization sharpens this line further. One formulation derives beampattern error, secrecy rate, and AoI as metrics for sensing, secrecy communication, and computing, then solves a multi-objective optimization problem with a DQN-based multi-objective evolutionary algorithm. A distinct defense line models channel-access attacks through a three-player Stackelberg game in which a malicious attacker is the leader and the legitimate drone and ground base station are followers. The backward-induction algorithm is reported to converge to a stable solution and to outperform existing baselines, thereby ensuring reliable ISAC performance for critical low-altitude applications (Wang et al., 3 Nov 2025, Wang et al., 9 Nov 2025).

Governance and trust are treated as architectural, not peripheral, functions. The low-altitude supervision system includes identity authentication, certificates, biometrics, aircraft codes, airworthiness databases, investigation, penalties, and public disclosure. Flight service systems perform task scheduling and inter-departmental resource sharing, while flight control systems optimize airspace capacity and issue real-time traffic commands (Wang et al., 30 Apr 2025). In a more economic formulation, aircraft computility is tokenized as Real-World Assets on a hybrid blockchain with permissioned and permissionless components; a fungible token called COMP represents units of computation, and smart contracts coordinate resource discovery, task auctions, payments, execution verification, and reputation. The reported simulations indicate improvements in task latency, trust assurance, and resource efficiency when leveraging RWA-based coordination (Luo et al., 25 Aug 2025).

Current research directions emphasize convergence rather than fragmentation. One line proposes an integrated air–ground computing and communication system with cloud–edge–air collaboration and an agentification paradigm, where a centralized orchestrator, edge agents, and flying agents interact through MCP and Duplex MCP for closed-loop re-planning (Sun et al., 24 Nov 2025). Another makes HAPs central to a global–regional–local three-tier architecture, assigning them regional orchestration, onboard computing and caching, navigation integrity, and airspace trust, with a five-stage roadmap extending from aerial infrastructure bases to edge–air–cloud closed-loop autonomy (Huang et al., 23 Feb 2026). Taken together, these works suggest that future LACNets will be increasingly digital-twin-enhanced, security-aware, explainability-constrained, and interoperable across cloud, edge, air, and space tiers, with communication, computation, and regulation treated as co-equal system primitives rather than separable subsystems.