UAV-Enabled Computing Power Network

• UAV-CPN is a paradigm that integrates UAVs as dynamic airborne agents into a four-tier cloud–edge–air–terminal architecture for real-time computing and sensing.
• It leverages hierarchical orchestration and dynamic resource allocation to manage latency-sensitive tasks while balancing energy and mobility constraints.
• Advanced orchestration techniques including DRL-based scheduling and distributed protocols significantly boost task success rates and network resilience.

A UAV-Enabled Computing Power Network (UAV-CPN) is a paradigm integrating unmanned aerial vehicles (UAVs) as mobile computing and communication agents within broader computing power networks (CPNs). UAV-CPNs exploit the mobility, sensing, and coverage of low-altitude LAE (Low-Altitude Economy) platforms to jointly deliver computing and communication services, leveraging synergy with terrestrial edge and cloud resources. These networks combine hierarchical orchestration, dynamic resource allocation, and multi-agent intelligence to optimize the execution of latency-sensitive, computationally intensive tasks over dynamic spatio-temporal demand patterns (Sun et al., 24 Nov 2025).

1. Architectural Foundations

A UAV-CPN fundamentally extends the traditional CPN topology from a three-tier Cloud–Edge–Terminal structure to a four-tier Cloud–Edge–Air–Terminal framework:

• Cloud Tier: Core data centers possess nearly unlimited compute/storage capacity ($f_C$), performing centralized orchestration, global model training, and system-wide resource management.
• Edge Tier: Fixed micro-data centers (%%%%1%%%% for node $i$) process latency-sensitive tasks, cache data, and instantiate VNFs (virtual network functions).
• Air Tier (UAVs): A fleet of $N$ UAVs, each with compute $f_{U,j}$, battery energy $E_j$, adjustable positions $p_j(t)$, and sensors. UAVs act as "Flying Agents"—relaying links, conducting aerial sensing, and enabling partial compute at locations inaccessible to terrestrial nodes.
• Terminal Tier: End-user devices generate data and task streams, dynamically offloading to the Air/Edge layers.

UAV-CPNs employ a centralized orchestrator (typically cloud-based) that interacts with both edge and air agents via the Model Context Protocol (MCP) stack, supporting duplex RPC for real-time planning and reflection. This closed-loop control facilitates adaptive Computing-as-a-Service that exploits all tiers for workload distribution. Agentification is central: each node—UAV or ground—acts as a perception–planning–action–reflection cycle agent, reporting its state, executing assignments, and triggering re-planning when significant deviations (e.g., battery depletion, link outage) occur (Sun et al., 24 Nov 2025).

2. Mathematical System Model

The standard UAV-CPN system is represented by the following entities and notation:

• Sets: $D = \{1,\dots,D\}$ (tasks), $J = \{1,\dots,N\}$ (UAVs), $I = \{1,\dots,K\}$ (edge nodes), $C$ (cloud).
• Task parameters: For $d\in D$, workload $L_d$ (cycles), input size $S_d$ (bits), latency bound $T_d$.
• Computing capacities: UAV $j$: $f_{U,j}$; Edge $i$: $f_{E,i}$.
• Communication rates: $r_{j\to i} = B\log_2(1 + P_j h_{j,i}/N_0)$; similar for other link types.
• Mobility & energy constraints: UAV position $p_j(t)\in\mathbb{R}^2$, velocity bound $\|\dot{p}_j(t)\|\le v_{\max}$, battery constraint $$E_{\text{fly},j} + E_{\text{comm},j} + E_{\text{comp},j} \le E_j$$.

Latency components for offloaded task $d$ to node $n$:

• Uplink: $t_{\text{up},d,n} = S_d / r_{\text{u}\to n}$
• Computation: $t_{\text{comp},d,n} = L_d / f_n$
• Downlink: $t_{\text{down},d,n} = O_d / r_{n\to\text{u}}$ ($O_d$ is output size)

Energy models:

• Computation: $$E_{\text{comp},n}^{(d)} = \kappa_n \cdot L_d \cdot (f_n)^2$$
• Transmission: $$E_{\text{tx},j}^{(d)} = P_j^{\text{tx}} \cdot (S_d / r_{j\to i})$$

Decision variables comprise binary offloading assignments ($x_{d,n}\in\{0,1\}$), UAV trajectory variables ($y_j(t)$), and routing variables ($z_{d,l}$).

3. Optimization and Orchestration

The UAV-CPN global objective is a weighted minimization of cumulative latency and energy across all tasks and computing nodes: $$\min_{x,y,z}\; \sum_{d\in D}\sum_{n}x_{d,n}\left[ \alpha\, (t_{\text{up},d,n} + t_{\text{comp},d,n} + t_{\text{down},d,n}) + \beta\, (E_{\text{tx},d,n} + E_{\text{comp},d,n}) \right]$$ Subject to latency, computation capacity, energy budget, and mobility constraints (Sun et al., 24 Nov 2025).

This formulation is typically a Mixed-Integer Nonlinear Program (MINLP). Standard solution approaches include Lagrangian dual decomposition, successive convex approximation, or heuristic Deep Reinforcement Learning (DRL) agents embedded as decision-makers in the orchestrator. DRL-based orchestrators learn to jointly schedule tasks, communication flows, and UAV mobility in response to global state observations (Sun et al., 24 Nov 2025).

4. Agentification and Hotspot Response

The agentification paradigm operationalizes each UAV and edge node as a multi-phase agent:

• Perception: Flying Agents sense the local environment, reporting link metrics, demand hotspots, and obstacles.
• Planning: Centralized orchestration aggregates agent states to optimize computing–communication scheduling and route/trajectory assignments.
• Action: Nodes execute their duties—instantiating VNFs, computing, relaying, or repositioning.
• Reflection: Duplex MCP ships real-time feedback to the orchestrator on execution anomalies, prompting plan adaptation.

Case studies demonstrate that during dynamic hotspot events, UAV-CPN agentification substantially outperforms both pure CPN (static edge only) and pure LAE (UAV only) architectures. For intense hotspot scenarios, integrated agentified CPN–LAE yields up to $\approx 85\%$ task success rate, compared to $45\%$ (CPN) and $30\%$ (LAE), maintaining near-optimal performance through rapid UAV repositioning and load redistribution (Sun et al., 24 Nov 2025).

5. Design Principles and Implementation Challenges

System design must balance communication and computation resources to maximize task completion probability under latency constraints. Key principles include:

• UAV altitude: There exists an optimal $h^*$ balancing LoS probability and path-loss, maximizing successful reachability of computing nodes (CNs) within latency budgets (Deng et al., 17 Dec 2025).
• CN density: Increasing CN density $\lambda_c$ enhances success probability, with diminishing returns as density saturates communication coverage (Deng et al., 17 Dec 2025).
• Hybrid energy: UAV propulsion (fuel cell) and communication (battery) dual-energy models require joint altitude–power adaptation; constraint violation reduces serviceability (Deng et al., 14 Jan 2026).
• Resource orchestration: Precompute feasibility maps for $(h,P_d)$ combinations and embed real-time iterative optimizers on-board for adaptation to changing GU/CN statistics.

Task completion probability, defined as the probability of finishing the task within $T_{\max}$, is analytically derived via stochastic geometry and Poisson thinning: $$P_{\rm success}(r_u) = 1 - \exp\Big\{ -2\pi \lambda_c \int_0^{r_c^{\max}(r_u)} F_{t_c}(T_{\max} - t_1(r_u) - t_2(r_c); D) r_c\, dr_c \Big\}$$ (Deng et al., 17 Dec 2025, Deng et al., 14 Jan 2026).

6. Distributed Protocols and Computation Paradigms

UAV-CPNs harness distributed consensus, coded computation, and privacy-preserving task offloading:

• Consensus/flocking protocols: Enable rapid formation, robustness to link failures, and resource sharing in UAV swarms (Wu et al., 2022).
• Dynamic coded computation: On-demand redundancy using MDS codes mitigates straggler nodes and ensures robust, privacy-aware distributed convolution, significantly accelerating latency and resilience compared to baseline uncoded methods (Zhou et al., 2022).
• Blockchain and decentralized ledgers: Allocate tamper-proof accounting for cooperative CPU usage and smart-contract scheduling of "compute-coins" to avoid freeloading (Wu et al., 2022).
• Federated/dist split AI training: UAVs collaboratively train models through federated updates or split learning, reducing transmission overhead and maintaining privacy (Sun et al., 28 Sep 2025).

7. Future Directions and Open Research Problems

Critical outstanding challenges for UAV-CPN design and deployment include:

• Power and Endurance: Developing energy harvesting mechanisms (solar, wireless) and advanced trajectory–compute optimization to prolong mission times.
• Link adaptation: Addressing rapid air–ground channel variations, improving link estimation, and deploying predictive rate/trajectory control.
• Multi-tenant orchestration: Ensuring service isolation, secure resource slicing, and fairness in shared air–ground architectures.
• Security: Mitigating threats of hijacking, spoofing, and eavesdropping via lightweight encryption, blockchain trust, and privacy-preserving computation (MPC, federated learning).
• AI-native orchestration: Enabling end-to-end DRL-based resource scheduling encompassing both compute and mobility in massive UAV swarms.
• Digital twins: Creating virtual replicas for predictive management, “what-if” simulations, and dynamic policy tuning.
• Scalability: Ensuring reliable, low-latency consensus, interference management, and task re-planning in large fleets, especially as network scale and density increase (Wu et al., 2022, Sun et al., 24 Nov 2025).

Ongoing research is exploring multi-tier integrated space–air–ground networks, leveraging adaptive offloading strategies and hierarchy-aware resource allocation to exploit satellite and terrestrial infrastructure for coverage and compute augmentation (Liu et al., 2024, Traspadini et al., 2023).

---

By architecting UAV-CPNs around integrated agent-based orchestration, energy-aware communication planning, and distributed computation, researchers and practitioners can substantially enhance system flexibility, responsiveness, and coverage, addressing key demands in emergent low-altitude services and high-density edge computing scenarios (Sun et al., 24 Nov 2025).