pith. sign in

arxiv: 2605.03898 · v1 · submitted 2026-05-05 · 📡 eess.SP

Joint Scheduling of Sensing Data Offloading and Edge Inference for Multi-UAV Networks

Yanan Du , Sai Xu , Yinbo Yu This is my paper

Pith reviewed 2026-05-07 14:00 UTC · model grok-4.3

classification 📡 eess.SP
keywords multi-UAV networksedge inferencesensing data offloadinggenetic algorithmend-to-end latencymulti-branch DNNsynchronization penaltyjoint scheduling
0
0 comments X

The pith

Genetic algorithm joint scheduling reduces end-to-end latency for multi-UAV sensing data offloading and edge inference.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper sets up a model where multiple UAVs collect sensing streams and offload them to an edge server running a multi-branch deep neural network on a multi-core accelerator. It formulates the problem of minimizing total latency including a penalty for synchronizing the different streams. The authors propose a genetic algorithm scheduler called GA-Joint for the full joint problem and two simplified versions GA-DAG and GA-DACS to cut computation time. Simulations indicate these approaches deliver lower latency than both decoupled greedy and joint greedy methods in most tested scenarios. Readers interested in real-time UAV applications would care because faster inference enables quicker decision-making from fused sensor data.

Core claim

A multi-UAV collaborative edge inference model is established where UAV sensing streams are processed by a multi-branch DNN on a multi-core accelerator. An end-to-end latency minimization problem with a synchronization penalty is formulated. A genetic algorithm-based full joint scheduler termed GA-Joint is developed, along with lightweight variants GA-DAG and GA-DACS. These achieve lower end-to-end latency than Decoupled-Greedy and Joint-Greedy in simulations.

What carries the argument

The genetic algorithm joint scheduler (GA-Joint) and its GA-DAG and GA-DACS variants that optimize the coupled decisions of data offloading from UAVs and multi-branch DNN execution on the edge accelerator.

Load-bearing premise

The assumption that wireless offloading times are deterministic or predictable and multi-branch DNN execution times on the multi-core accelerator are accurately known in advance.

What would settle it

A real-world experiment deploying the GA schedulers on physical UAVs and an edge server, then comparing measured end-to-end latency against the Decoupled-Greedy and Joint-Greedy baselines.

Figures

Figures reproduced from arXiv: 2605.03898 by Sai Xu, Yanan Du, Yinbo Yu.

Figure 1
Figure 1. Figure 1: An illustration of the multi-UAV collaborative edge inference system. view at source ↗
Figure 2
Figure 2. Figure 2: Convergence behavior of the three GA-based scheduling schemes. view at source ↗
Figure 3
Figure 3. Figure 3: Comparison of end-to-end execution timelines under the evaluated scheduling schemes. view at source ↗
Figure 6
Figure 6. Figure 6: Impact of the SINR threshold on end-to-end latency. view at source ↗
Figure 5
Figure 5. Figure 5: Impact of the number of subcarriers on end-to-end latency. view at source ↗
read the original abstract

Unmanned aerial vehicles (UAVs) often collaborate by collecting and offloading sensing streams to an edge server, where a deep neural network (DNN) model performs cross-stream alignment, fusion, and inference. However, the coupling between wireless offloading and DNN execution makes end-to-end latency minimization challenging. To address this issue, this paper investigates efficient edge inference in multi-UAV networks. Specifically, a multi-UAV collaborative edge inference model is first established, in which UAV sensing streams are processed by a multi-branch DNN on a multi-core accelerator. Based on this model, an end-to-end latency minimization problem with a synchronization penalty is formulated. A genetic algorithm (GA)-based full joint scheduler, termed \texttt{GA-Joint}, is then developed to obtain high-quality scheduling solutions. To reduce the search complexity, two lightweight variants, termed \texttt{GA-DAG} and \texttt{GA-DACS}, are further proposed. Simulation results demonstrate that the proposed GA-based scheduling algorithms achieve lower end-to-end latency than \texttt{Decoupled-Greedy} and \texttt{Joint-Greedy}, which represent decoupled and joint greedy scheduling schemes, respectively, in most cases. Furthermore, \texttt{GA-DACS} achieves performance close to that of \texttt{GA-Joint} in many cases and even delivers slightly lower latency in certain scenarios.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 2 minor

Summary. The paper models multi-UAV collaborative edge inference with a multi-branch DNN executed on a multi-core accelerator, formulates an end-to-end latency minimization problem that includes a synchronization penalty, and develops three GA-based schedulers (GA-Joint, GA-DAG, GA-DACS) whose performance is evaluated via simulation against Decoupled-Greedy and Joint-Greedy baselines. The central claim is that the proposed GA variants achieve lower latency than the greedy schemes in most simulated cases, with GA-DACS often close to GA-Joint.

Significance. If the simulation results hold under the stated modeling assumptions, the work provides concrete heuristic schedulers for the coupled offloading-plus-inference problem in multi-UAV edge networks. The explicit formulation of the joint optimization and the introduction of two reduced-complexity GA variants constitute the main technical contribution; the simulation evidence of latency reduction is the primary empirical support.

major comments (2)
  1. [Simulation Results] Simulation Results section: the claim that the GA schedulers achieve lower end-to-end latency “in most cases” is load-bearing for the paper’s contribution, yet the manuscript provides no details on the number of Monte-Carlo runs, channel model parameters, UAV mobility traces, or statistical significance tests (e.g., confidence intervals or p-values). Without these, it is impossible to assess whether the observed gains are robust or sensitive to the chosen deterministic offloading and DNN timing assumptions.
  2. [System Model] System Model and Problem Formulation sections: the multi-branch DNN execution time on the multi-core accelerator and the synchronization penalty are treated as deterministic and perfectly known; the paper does not discuss sensitivity of the GA solutions to errors in these quantities or to stochastic wireless channel realizations, which directly affects whether the latency-minimization claim translates beyond the simulated instances.
minor comments (2)
  1. [Abstract] Abstract: the simulation parameters, channel models, and number of trials are not mentioned, making it difficult for readers to gauge the scope of the reported latency improvements.
  2. [Proposed Algorithms] Notation: the distinction between the three GA variants (GA-Joint, GA-DAG, GA-DACS) is introduced only in the abstract and algorithm descriptions; a compact table summarizing their search spaces and complexity would improve readability.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive comments, which help strengthen the clarity and robustness of our work. We address each major comment below and will revise the manuscript to incorporate the suggested improvements where feasible.

read point-by-point responses

  1. Referee: [Simulation Results] Simulation Results section: the claim that the GA schedulers achieve lower end-to-end latency “in most cases” is load-bearing for the paper’s contribution, yet the manuscript provides no details on the number of Monte-Carlo runs, channel model parameters, UAV mobility traces, or statistical significance tests (e.g., confidence intervals or p-values). Without these, it is impossible to assess whether the observed gains are robust or sensitive to the chosen deterministic offloading and DNN timing assumptions.

    Authors: We agree that the simulation setup requires more explicit documentation to support the performance claims. In the revised manuscript, we will expand the Simulation Results section to specify the number of Monte-Carlo runs (500 independent trials per scenario), the full channel model parameters (including path-loss exponent, shadowing variance, and Rician fading factors), the UAV mobility model (random waypoint with maximum speed of 20 m/s and pause times), and statistical measures such as 95% confidence intervals on the reported latency values. These additions will enable readers to evaluate the robustness of the latency reductions relative to the greedy baselines. revision: yes

  2. Referee: [System Model] System Model and Problem Formulation sections: the multi-branch DNN execution time on the multi-core accelerator and the synchronization penalty are treated as deterministic and perfectly known; the paper does not discuss sensitivity of the GA solutions to errors in these quantities or to stochastic wireless channel realizations, which directly affects whether the latency-minimization claim translates beyond the simulated instances.

    Authors: The formulation adopts deterministic values for DNN execution times and synchronization penalties to maintain tractability of the joint optimization problem. We acknowledge that this limits direct applicability to stochastic environments. In the revision, we will insert a dedicated paragraph in the System Model section that analyzes sensitivity of the GA schedulers to ±10% perturbations in DNN timing and synchronization estimates, and we will add a brief discussion of how channel variability could be incorporated via robust or stochastic variants of the GA. We will also clarify that the current results hold under the stated modeling assumptions and flag stochastic extensions as future work. revision: partial

Circularity Check

0 steps flagged

No significant circularity

full rationale

The paper establishes a multi-UAV collaborative edge inference model, formulates an explicit end-to-end latency minimization problem with synchronization penalty, and solves it via external genetic algorithms (GA-Joint and variants) whose outputs are compared empirically against greedy baselines in simulation. No load-bearing step reduces a claimed result to a fitted parameter, self-defined quantity, or self-citation chain by construction; the performance claims are direct simulation outcomes on the stated model rather than derivations that presuppose their own outputs.

Axiom & Free-Parameter Ledger

0 free parameters · 2 axioms · 0 invented entities

The central claim rests on a domain-specific model of multi-branch DNN execution and wireless offloading that is not derived from first principles but taken as given for the optimization.

axioms (2)
  • domain assumption Multi-branch DNN execution times on a multi-core accelerator and wireless offloading durations can be modeled with sufficient accuracy for end-to-end latency minimization.
    Invoked when formulating the optimization problem and evaluating the GA solutions.
  • domain assumption The synchronization penalty term correctly captures the cost of waiting for all sensing streams to arrive before fusion and inference.
    Central to the end-to-end latency objective.

pith-pipeline@v0.9.0 · 5549 in / 1280 out tokens · 63797 ms · 2026-05-07T14:00:51.039626+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

29 extracted references · 29 canonical work pages

  1. [1]

    MUL-VR: Multi-UA V collaborative layered visual perception and transmission for virtual reality,

    X.-W. Tang, Y . Huang, Y . Shi, and Q. Wu, “MUL-VR: Multi-UA V collaborative layered visual perception and transmission for virtual reality,”IEEE Trans. Wireless Commun., vol. 24, no. 4, pp. 2734–2749, 14 Apr. 2025

    work page 2025

  2. [2]

    UCDNet: Multi-UA V collaborative 3-D object detection network by reliable feature mapping,

    P. Tianet al., “UCDNet: Multi-UA V collaborative 3-D object detection network by reliable feature mapping,”IEEE Trans. Geosci. Remote Sens., vol. 63, pp. 1–16, 2025, Art. no. 5602016

    work page 2025

  3. [3]

    U2UData: A large-scale cooperative perception dataset for swarm UA Vs autonomous flight,

    T. Feng, X. Wang, F. Han, L. Zhang, and W. Zhu, “U2UData: A large-scale cooperative perception dataset for swarm UA Vs autonomous flight,” inProc. ACM Int. Conf. Multimedia (ACM MM), 2024, pp. 7600– 7608

    work page 2024

  4. [4]

    UA VScenes: A multi-modal dataset for UA Vs,

    S. Wanget al., “UA VScenes: A multi-modal dataset for UA Vs,” inProc. IEEE/CVF Int. Conf. Comput. Vis. (ICCV), 2025, pp. 28946–28958

    work page 2025

  5. [5]

    Resource al- location for multi-modal semantic communication in UA V collaborative networks,

    H. Hu, X. Zhu, F. Zhou, W. Wu, R. Q. Hu, and H. Zhu, “Resource al- location for multi-modal semantic communication in UA V collaborative networks,”IEEE Trans. Commun., vol. 73, no. 9, pp. 7599–7616, Sep. 2025

    work page 2025

  6. [6]

    Differential multimodal fusion algorithm for remote sensing object detection,

    W. Zhaoet al., “Differential multimodal fusion algorithm for remote sensing object detection,”Expert Syst. Appl., vol. 261, 2025, Art. no. 125485

    work page 2025

  7. [7]

    UA V-based multimodal object detection via feature enhancement and dynamic gated fusion,

    Y . Gu, W. Chen, and D. Peng, “UA V-based multimodal object detection via feature enhancement and dynamic gated fusion,”Pattern Recognit., vol. 169, 2026, Art. no. 111930

    work page 2026

  8. [8]

    UA V-enabled multi-tier mobile edge computing for heterogeneous dual-source multi-modal tasks,

    X. Youet al., “UA V-enabled multi-tier mobile edge computing for heterogeneous dual-source multi-modal tasks,”IEEE Wireless Commun. Lett., early access, 2026, doi: 10.1109/LWC.2026.3684964

    work page doi:10.1109/lwc.2026.3684964 2026

  9. [9]

    Offloading deep learning powered vision tasks from UA V to 5G edge server with denoising,

    S. Ozer, H. E. Ilhan, M. A. Ozkanoglu, and H. A. Cirpan, “Offloading deep learning powered vision tasks from UA V to 5G edge server with denoising,”IEEE Trans. Veh. Technol., vol. 72, no. 6, pp. 8035–8048, Jun. 2023

    work page 2023

  10. [10]

    MEC-assisted real-time data acquisition and processing for UA V with general missions,

    Y . Zeng and J. Tang, “MEC-assisted real-time data acquisition and processing for UA V with general missions,”IEEE Trans. Veh. Technol., vol. 72, no. 1, pp. 1058–1072, Jan. 2023

    work page 2023

  11. [11]

    Blockchain-based MIMO UA V-aided mobile edge computing,

    X. Donget al., “Blockchain-based MIMO UA V-aided mobile edge computing,”IEEE Trans. Mobile Comput., early access, 2025, doi: 10.1109/TMC.2025.3649700

    work page doi:10.1109/tmc.2025.3649700 2025

  12. [12]

    Hybrid beamforming design and resource allocation for UA V-aided wireless-powered mobile edge computing networks with NOMA,

    W. Fenget al., “Hybrid beamforming design and resource allocation for UA V-aided wireless-powered mobile edge computing networks with NOMA,”IEEE J. Sel. Areas Commun., vol. 39, no. 11, pp. 3271–3286, Nov. 2021

    work page 2021

  13. [13]

    MAESTRO: A data-centric approach to understand reuse, performance, and hardware cost of DNN mappings,

    H. Kwon, P. Chatarasi, V . Sarkar, T. Krishna, M. Pellauer, and A. Parashar, “MAESTRO: A data-centric approach to understand reuse, performance, and hardware cost of DNN mappings,”IEEE Micro, vol. 40, no. 3, pp. 20–29, May–Jun. 2020

    work page 2020

  14. [14]

    Magma: An optimization framework for mapping multiple DNNs on multiple accelerator cores,

    S.-C. Kao and T. Krishna, “Magma: An optimization framework for mapping multiple DNNs on multiple accelerator cores,” inProc. IEEE Int. Symp. High-Performance Comput. Archit. (HPCA), 2022, pp. 814– 830

    work page 2022

  15. [15]

    Edge AI: On-Demand acceler- ating deep neural network inference via edge computing,

    E. Li, L. Zeng, Z. Zhou, and X. Chen, “Edge AI: On-Demand acceler- ating deep neural network inference via edge computing,”IEEE Trans. Wireless Commun., vol. 19, no. 1, pp. 447–457, Jan. 2020

    work page 2020

  16. [16]

    Energy-efficient processing and robust wireless cooperative transmission for edge inference,

    K. Yang, Y . Shi, W. Yu, and Z. Ding, “Energy-efficient processing and robust wireless cooperative transmission for edge inference,”IEEE Internet Things J., vol. 7, no. 10, pp. 9456–9470, Oct. 2020

    work page 2020

  17. [17]

    Throughput maximization of delay-aware DNN inference in edge computing by exploring DNN model partitioning and inference parallelism,

    J. Li, W. Liang, Y . Li, Z. Xu, X. Jia, and S. Guo, “Throughput maximization of delay-aware DNN inference in edge computing by exploring DNN model partitioning and inference parallelism,”IEEE Trans. Mobile Comput., vol. 22, no. 5, pp. 3017–3030, May 2023

    work page 2023

  18. [18]

    Learning task-oriented communication for edge inference: An information bottleneck approach,

    J. Shao, Y . Mao, and J. Zhang, “Learning task-oriented communication for edge inference: An information bottleneck approach,”IEEE J. Sel. Areas Commun., vol. 40, no. 1, pp. 197–211, Jan. 2022

    work page 2022

  19. [19]

    FrankenSplit: Efficient neural feature compression with shallow variational bottleneck injection for mobile edge computing,

    A. Furutanpey, P. Raith, and S. Dustdar, “FrankenSplit: Efficient neural feature compression with shallow variational bottleneck injection for mobile edge computing,”IEEE Trans. Mobile Comput., vol. 23, no. 12, pp. 10770–10786, Dec. 2024

    work page 2024

  20. [20]

    Tackling distribution shifts in task-oriented communication with information bottleneck,

    H. Li, J. Shao, H. He, S. Song, J. Zhang, and K. B. Letaief, “Tackling distribution shifts in task-oriented communication with information bottleneck,”IEEE J. Sel. Areas Commun., vol. 43, no. 7, pp. 2667– 2683, Jul. 2025

    work page 2025

  21. [21]

    Adaptable variational information bottleneck for task-oriented edge inference,

    E. Tarimo, H. Xing, L. Xu, J. Peng, and L. Feng, “Adaptable variational information bottleneck for task-oriented edge inference,”IEEE Trans. Netw. Sci. Eng., vol. 13, pp. 8574–8592, 2026

    work page 2026

  22. [22]

    Toward real- time edge AI: Model-agnostic task-oriented communication with visual feature alignment,

    S. Xie, H. He, S. Song, J. Zhang, and K. B. Letaief, “Toward real- time edge AI: Model-agnostic task-oriented communication with visual feature alignment,”IEEE J. Sel. Areas Commun., vol. 43, no. 12, pp. 4262–4276, Dec. 2025

    work page 2025

  23. [23]

    A multi-neural network acceleration architecture,

    E. Baek, D. Kwon, and J. Kim, “A multi-neural network acceleration architecture,” inProc. ACM/IEEE 47th Annu. Int. Symp. Comput. Archit. (ISCA), 2020, pp. 940–953

    work page 2020

  24. [24]

    Memory and computation coordi- nated mapping of DNNs onto complex heterogeneous SoC,

    S. Zheng, S. Chen, and Y . Liang, “Memory and computation coordi- nated mapping of DNNs onto complex heterogeneous SoC,” inProc. ACM/IEEE 60th Design Autom. Conf. (DAC), 2023, pp. 1–6

    work page 2023

  25. [25]

    MoCA: Memory-centric, adaptive execution for multi-tenant deep neural networks,

    S. Kim, H. Genc, V . V . Nikiforov, K. Asanovi ´c, B. Nikoli ´c, and Y . S. Shao, “MoCA: Memory-centric, adaptive execution for multi-tenant deep neural networks,” inProc. IEEE Int. Symp. High-Performance Comput. Archit. (HPCA), 2023, pp. 828–841

    work page 2023

  26. [26]

    Heterogeneous dataflow accelerators for multi-DNN workloads,

    H. Kwon, L. Lai, M. Pellauer, T. Krishna, Y .-H. Chen, and V . Chandra, “Heterogeneous dataflow accelerators for multi-DNN workloads,” in Proc. IEEE Int. Symp. High-Performance Comput. Archit. (HPCA), 2021, pp. 71–83

    work page 2021

  27. [27]

    DREAM: A dynamic scheduler for dynamic real-time multi-model ML workloads,

    S. Kim, H. Kwon, J. Song, J. Jo, Y .-H. Chen, L. Lai, and V . Chandra, “DREAM: A dynamic scheduler for dynamic real-time multi-model ML workloads,” inProc. 28th ACM Int. Conf. Archit. Support Program. Lang. Oper. Syst. (ASPLOS), vol. 4, 2023, pp. 73–86

    work page 2023

  28. [28]

    Sparse-DySta: Sparsity- aware dynamic and static scheduling for sparse multi-DNN workloads,

    H. Fan, S. I. Venieris, A. Kouris, and N. Lane, “Sparse-DySta: Sparsity- aware dynamic and static scheduling for sparse multi-DNN workloads,” inProc. 56th Annu. IEEE/ACM Int. Symp. Microarchitecture (MICRO), 2023, pp. 353–366

    work page 2023

  29. [29]

    TaiChi: Efficient execution for multi-DNNs using graph- based scheduling,

    X. Zhouet al., “TaiChi: Efficient execution for multi-DNNs using graph- based scheduling,” inProc. Design, Autom. Test Europe Conf. Exhib. (DATE), Lyon, France, 2025, pp. 1–7

    work page 2025