pith. sign in

arxiv: 2406.13086 · v2 · submitted 2024-06-18 · 💻 cs.RO · cs.AI

NaviSplit: Dynamic Multi-Branch Split DNNs for Efficient Distributed Autonomous Navigation

Timothy K Johnsen , Ian Harshbarger , Zixia Xia , Marco Levorato This is my paper

Pith reviewed 2026-05-23 23:33 UTC · model grok-4.3

classification 💻 cs.RO cs.AI
keywords split DNNneural gateUAV navigationdistributed inferencedepth estimationdynamic multi-branchedge computingautonomous navigation
0
0 comments X

The pith

A neural gate dynamically selects among split DNN heads on a UAV to deliver 0.3 percent higher navigation accuracy than a static network while cutting transmitted data by 95 percent.

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

The paper presents NaviSplit as a framework that splits deep neural networks for perception and control across a resource-limited UAV and a connected edge server. A head model runs on the vehicle to partially process camera images and compress the output, while the tail model on the server completes the computation to produce navigation commands. A neural gate chooses among several head models at runtime to keep the transmitted data volume small. Depth extraction from monocular RGB images reaches 72 to 81 percent accuracy while sending only 1.2 to 18 kilobytes per frame. The gated dynamic version improves overall navigation accuracy by 0.3 percent over a single larger static network and reduces the data rate by 95 percent.

Core claim

NaviSplit is the first dynamic multi-branched split DNN framework for autonomous UAV navigation. The network is divided at a compression point so the vehicle executes one of several head models that compact sensor perception, and an edge server executes the tail model that produces navigation commands. The neural gate selects the active head to minimize channel usage. In Microsoft AirSim experiments the depth model extracts maps at 72-81 percent accuracy while transmitting 1.2-18 KB, and the full gated system yields 0.3 percent higher navigation accuracy than a larger static network together with a 95 percent reduction in data rate.

What carries the argument

The neural gate, which dynamically selects among multiple head models to minimize transmitted data volume while preserving navigation performance.

If this is right

  • The vehicle-side head model compacts perception data before transmission to the edge server.
  • The edge-side tail model completes inference to generate navigation commands from the compacted input.
  • Depth extraction accuracy stays between 72 and 81 percent while sending between 1.2 and 18 KB.
  • The neural gate enables the system to exceed the navigation accuracy of a single larger static network.
  • Data transmission rate drops by 95 percent relative to the static baseline without loss of accuracy.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The same split-and-gate pattern could be applied to other onboard perception tasks such as semantic segmentation or obstacle avoidance.
  • Variable wireless conditions in real deployments would likely require the gate to also consider instantaneous channel quality when choosing heads.
  • Because branches differ in size, the approach may allow graceful degradation when the edge link is temporarily unavailable.

Load-bearing premise

The Microsoft AirSim simulator produces depth and navigation behavior sufficiently representative of real-world UAV flight that the reported accuracy and data-rate gains will translate outside simulation.

What would settle it

Deploying the NaviSplit models on physical UAVs in outdoor flight and comparing measured navigation accuracy and bytes transmitted per frame against the AirSim results.

Figures

Figures reproduced from arXiv: 2406.13086 by Ian Harshbarger, Marco Levorato, Timothy K Johnsen, Zixia Xia.

Figure 1
Figure 1. Figure 1: The framework we propose uses a teacher model that maps a [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Comparing various compressed data sizes, corresponding to different [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Mean value of the gate control, cˆ, for all successful paths while using an auxiliary model to adapt and select cˆ at each time step. This a demonstration of NaviSplit, which was trained and evaluated in Microsoft AirSim. . V. CONCLUSIONS We have presented NaviSplit – an effective multi-branched neural network architecture for drone navigation, that dy￾namically splits computing between an equipped process… view at source ↗
read the original abstract

Lightweight autonomous unmanned aerial vehicles (UAV) are emerging as a central component of a broad range of applications. However, autonomous navigation necessitates the implementation of perception algorithms, often deep neural networks (DNN), that process the input of sensor observations, such as that from cameras and LiDARs, for control logic. The complexity of such algorithms clashes with the severe constraints of these devices in terms of computing power, energy, memory, and execution time. In this paper, we propose NaviSplit, the first instance of a lightweight navigation framework embedding a distributed and dynamic multi-branched neural model. At its core is a DNN split at a compression point, resulting in two model parts: (1) the head model, that is executed at the vehicle, which partially processes and compacts perception from sensors; and (2) the tail model, that is executed at an interconnected compute-capable device, which processes the remainder of the compacted perception and infers navigation commands. Different from prior work, the NaviSplit framework includes a neural gate that dynamically selects a specific head model to minimize channel usage while efficiently supporting the navigation network. In our implementation, the perception model extracts a 2D depth map from a monocular RGB image captured by the drone using the robust simulator Microsoft AirSim. Our results demonstrate that the NaviSplit depth model achieves an extraction accuracy of 72-81% while transmitting an extremely small amount of data (1.2-18 KB) to the edge server. When using the neural gate, as utilized by NaviSplit, we obtain a slightly higher navigation accuracy as compared to a larger static network by 0.3% while significantly reducing the data rate by 95%. To the best of our knowledge, this is the first exemplar of dynamic multi-branched model based on split DNNs for autonomous navigation.

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 proposes NaviSplit, the first dynamic multi-branch split DNN framework for distributed autonomous UAV navigation. A perception DNN is split at a compression point into a vehicle-side head (RGB-to-depth extraction and compaction) and an edge-side tail (navigation command inference), with an added neural gate that dynamically selects among multiple head branches to minimize transmitted data volume. All evaluation is performed inside the Microsoft AirSim simulator; the manuscript reports 72-81% depth extraction accuracy with 1.2-18 KB transmissions per frame and, when the neural gate is active, a 0.3% navigation accuracy improvement over a larger static network together with a 95% data-rate reduction.

Significance. If the reported accuracy and compression figures prove robust, the combination of model splitting with a learned dynamic gate would constitute a practical advance for communication-constrained robotic perception pipelines. The work supplies concrete empirical numbers on data volume and accuracy trade-offs inside a widely used simulator, which is a positive attribute for an applied systems paper.

major comments (2)
  1. [Abstract / Evaluation] Abstract and Evaluation section: the headline claims (72-81% depth accuracy, 0.3% navigation gain, 95% data-rate reduction) are stated without any description of experimental protocol, baseline network architectures and sizes, number of simulation episodes, random seeds, error bars, or statistical tests. Because these quantities are the sole support for the central performance assertions, the absence of this information is load-bearing.
  2. [Evaluation] Evaluation section: every quantitative result is obtained exclusively inside Microsoft AirSim. No real-world UAV flights, domain-randomization experiments, or sensitivity analysis to sensor noise, lighting variation, or wind disturbances are reported. The 0.3% accuracy delta and 95% compression benefit are therefore conditional on the untested assumption that AirSim dynamics and depth rendering are sufficiently representative of physical flight.
minor comments (2)
  1. [Related Work] The manuscript would benefit from an explicit comparison table placing NaviSplit against prior split-DNN and dynamic-exit works (e.g., BranchyNet, Neurosurgeon) in terms of split point, gate mechanism, and reported metrics.
  2. [Method] Notation for the neural gate (input features, output probabilities, training loss) is introduced only informally; a short equation or pseudocode block would improve reproducibility.

Simulated Author's Rebuttal

2 responses · 1 unresolved

We thank the referee for the constructive feedback on the manuscript. We address each major comment below and indicate planned revisions.

read point-by-point responses

  1. Referee: [Abstract / Evaluation] Abstract and Evaluation section: the headline claims (72-81% depth accuracy, 0.3% navigation gain, 95% data-rate reduction) are stated without any description of experimental protocol, baseline network architectures and sizes, number of simulation episodes, random seeds, error bars, or statistical tests. Because these quantities are the sole support for the central performance assertions, the absence of this information is load-bearing.

    Authors: We agree that additional experimental details are needed to support the reported figures. In the revised manuscript we will expand the Evaluation section with a full description of the experimental protocol, including baseline network architectures and parameter counts, the number of simulation episodes, random seeds, and any error bars or statistical tests applied. revision: yes

  2. Referee: [Evaluation] Evaluation section: every quantitative result is obtained exclusively inside Microsoft AirSim. No real-world UAV flights, domain-randomization experiments, or sensitivity analysis to sensor noise, lighting variation, or wind disturbances are reported. The 0.3% accuracy delta and 95% compression benefit are therefore conditional on the untested assumption that AirSim dynamics and depth rendering are sufficiently representative of physical flight.

    Authors: We acknowledge that all results are simulator-based. AirSim is a standard benchmark for UAV navigation, yet we recognize the limitation. In revision we will add an explicit limitations subsection discussing simulator assumptions and the challenges of sim-to-real transfer. We cannot add physical flight data at this stage. revision: partial

standing simulated objections not resolved
  • Real-world UAV flight experiments to validate the reported accuracy and compression figures

Circularity Check

0 steps flagged

No circularity; results are direct empirical measurements

full rationale

The paper introduces NaviSplit as a new framework and reports all key outcomes (depth extraction accuracy 72-81%, data volumes 1.2-18 KB, +0.3% navigation accuracy, 95% data-rate reduction) as measured quantities obtained from runs inside Microsoft AirSim. No equations, derivations, fitted parameters renamed as predictions, or self-citation load-bearing steps appear. The central claims rest on simulator experiments rather than any reduction of outputs to inputs by construction, satisfying the self-contained criterion.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

Abstract-only review yields no identifiable free parameters, axioms, or invented entities; the work is presented as an empirical system design rather than a derivation from first principles.

pith-pipeline@v0.9.0 · 5886 in / 1252 out tokens · 26363 ms · 2026-05-23T23:33:45.430012+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

28 extracted references · 28 canonical work pages · 3 internal anchors

  1. [1]

    Qonnx: Representing arbitrary-precision quantized neural networks,

    A. Pappalardo, Y . Umuroglu, M. Blott, J. Mitrevski, B. Hawks, N. Tran, V . Loncar, S. Summers, H. Borras, J. Muhizi, M. Trahms, S.-C. Hsu, S. Hauck, and J. Duarte, “Qonnx: Representing arbitrary-precision quantized neural networks,” 2022

    work page 2022

  2. [2]

    Scalable methods for 8-bit training of neural networks,

    R. Banner, I. Hubara, E. Hoffer, and D. Soudry, “Scalable methods for 8-bit training of neural networks,” 2018

    work page 2018

  3. [3]

    Training deep neural networks with low precision multiplications,

    M. Courbariaux, Y . Bengio, and J.-P. David, “Training deep neural networks with low precision multiplications,” 2015

    work page 2015

  4. [4]

    Distilling the Knowledge in a Neural Network

    G. Hinton, O. Vinyals, and J. Dean, “Distilling the knowledge in a neural network,” arXiv preprint arXiv:1503.02531 , 2015

    work page internal anchor Pith review Pith/arXiv arXiv 2015

  5. [5]

    Variational information distillation for knowledge transfer,

    S. Ahn, S. X. Hu, A. Damianou, N. D. Lawrence, and Z. Dai, “Variational information distillation for knowledge transfer,” 2019

    work page 2019

  6. [6]

    Apprentice: Using knowledge distillation techniques to improve low-precision network accuracy,

    A. Mishra and D. Marr, “Apprentice: Using knowledge distillation techniques to improve low-precision network accuracy,” 2017

    work page 2017

  7. [7]

    Squeezenext: Hardware-aware neural network design,

    A. Gholami, K. Kwon, B. Wu, Z. Tai, X. Yue, P. Jin, S. Zhao, and K. Keutzer, “Squeezenext: Hardware-aware neural network design,” 2018

    work page 2018

  8. [8]

    Solving critical events through mobile edge computing: An approach for smart cities,

    M. Sapienza, E. Guardo, M. Cavallo, G. Torre, G. Leombruno, and O. Tomarchio, “Solving critical events through mobile edge computing: An approach for smart cities,” 05 2016, pp. 1–5

    work page 2016

  9. [9]

    Videoedge: Processing camera streams using hierarchical clusters,

    C.-C. Hung, G. Ananthanarayanan, P. Bodik, L. Golubchik, M. Yu, P. Bahl, and M. Philipose, “Videoedge: Processing camera streams using hierarchical clusters,” in 2018 IEEE/ACM Symposium on Edge Computing (SEC) , 2018, pp. 115–131

    work page 2018

  10. [10]

    Seremas: Self-resilient mobile autonomoussystems through predictive edge computing,

    D. Callegaro, M. Levorato, and F. Restuccia, “Seremas: Self-resilient mobile autonomoussystems through predictive edge computing,” CoRR, vol. abs/2105.15105, 2021. [Online]. Available: https://arxiv.org/abs/ 2105.15105

    work page arXiv 2021

  11. [11]

    Split computing and early exiting for deep learning applications: Survey and research challenges,

    Y . Matsubara, M. Levorato, and F. Restuccia, “Split computing and early exiting for deep learning applications: Survey and research challenges,” ACM Computing Surveys , vol. 55, no. 5, pp. 1–30, 2022

    work page 2022

  12. [12]

    Bottlefit: Learning compressed representations in deep neural networks for effective and efficient split computing,

    Y . Matsubara, D. Callegaro, S. Singh, M. Levorato, and F. Restuccia, “Bottlefit: Learning compressed representations in deep neural networks for effective and efficient split computing,” in 2022 IEEE 23rd Inter- national Symposium on a World of Wireless, Mobile and Multimedia Networks (WoWMoM). IEEE, 2022, pp. 337–346

    work page 2022

  13. [13]

    Airsim: High-fidelity visual and physical simulation for autonomous vehicles,

    S. Shah, D. Dey, C. Lovett, and A. Kapoor, “Airsim: High-fidelity visual and physical simulation for autonomous vehicles,” in Field and Service Robotics: Results of the 11th International Conference . Springer, 2018, pp. 621–635

    work page 2018

  14. [14]

    Edge computing-assisted dnn image recognition system with progressive image retransmission,

    M. Nakahara, M. Nishimura, Y . Ushiku, T. Nishio, K. Maruta, Y . Nakayama, and D. Hisano, “Edge computing-assisted dnn image recognition system with progressive image retransmission,” IEEE Ac- cess, vol. 10, pp. 91 253–91 262, 2022

    work page 2022

  15. [15]

    Neurosurgeon: Collaborative intelligence between the cloud and mobile edge,

    Y . Kang, J. Hauswald, C. Gao, A. Rovinski, T. Mudge, J. Mars, and L. Tang, “Neurosurgeon: Collaborative intelligence between the cloud and mobile edge,” in Proceedings of the Twenty-Second International Conference on Architectural Support for Programming Languages and Operating Systems , ser. ASPLOS ’17. New York, NY , USA: Association for Computing Mach...

    work page doi:10.1145/3037697.3037698 2017

  16. [16]

    Slimmable encoders for flexible split dnns in bandwidth and resource constrained iot systems,

    J. S. Assine, E. Valle, M. Levorato et al. , “Slimmable encoders for flexible split dnns in bandwidth and resource constrained iot systems,” arXiv preprint arXiv:2306.12691 , 2023

    work page arXiv 2023

  17. [17]

    Group normalization,

    Y . Wu and K. He, “Group normalization,” in Computer Vision – ECCV 2018, V . Ferrari, M. Hebert, C. Sminchisescu, and Y . Weiss, Eds. Cham: Springer International Publishing, 2018, pp. 3–19

    work page 2018

  18. [18]

    Self- normalizing neural networks,

    G. Klambauer, T. Unterthiner, A. Mayr, and S. Hochreiter, “Self- normalizing neural networks,” in Proceedings of the 31st international conference on neural information processing systems , 2017, pp. 972– 981

    work page 2017

  19. [19]

    Adam: A Method for Stochastic Optimization

    D. P. Kingma and J. Ba, “Adam: A method for stochastic optimization,” arXiv preprint arXiv:1412.6980 , 2014

    work page internal anchor Pith review Pith/arXiv arXiv 2014

  20. [20]

    Early stopping-but when?

    L. Prechelt, “Early stopping-but when?” in Neural Networks: Tricks of the trade . Springer, 2002, pp. 55–69

    work page 2002

  21. [21]

    Deep Learning using Rectified Linear Units (ReLU)

    A. F. Agarap, “Deep learning using rectified linear units (relu),” arXiv preprint arXiv:1803.08375, 2018

    work page internal anchor Pith review Pith/arXiv arXiv 2018

  22. [22]

    Addressing function approxi- mation error in actor-critic methods,

    S. Fujimoto, H. Hoof, and D. Meger, “Addressing function approxi- mation error in actor-critic methods,” in International conference on machine learning . PMLR, 2018, pp. 1587–1596

    work page 2018

  23. [23]

    Stable-baselines3: Reliable reinforcement learning implementa- tions,

    A. Raffin, A. Hill, A. Gleave, A. Kanervisto, M. Ernestus, and N. Dor- mann, “Stable-baselines3: Reliable reinforcement learning implementa- tions,” The Journal of Machine Learning Research , vol. 22, no. 1, pp. 12 348–12 355, 2021

    work page 2021

  24. [24]

    Automatic differentiation in pytorch,

    A. Paszke, S. Gross, S. Chintala, G. Chanan, E. Yang, Z. DeVito, Z. Lin, A. Desmaison, L. Antiga, and A. Lerer, “Automatic differentiation in pytorch,” 2017

    work page 2017

  25. [25]

    Airsim drone racing lab,

    R. Madaan, N. Gyde, S. Vemprala, M. Brown, K. Nagami, T. Taubner, E. Cristofalo, D. Scaramuzza, M. Schwager, and A. Kapoor, “Airsim drone racing lab,” in Neurips 2019 competition and demonstration track . PMLR, 2020, pp. 177–191

    work page 2019

  26. [26]

    Autonomous drone racing with deep reinforcement learning,

    Y . Song, M. Steinweg, E. Kaufmann, and D. Scaramuzza, “Autonomous drone racing with deep reinforcement learning,” in 2021 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS) . IEEE, 2021, pp. 1205–1212

    work page 2021

  27. [27]

    Air learning: a deep reinforcement learning gym for autonomous aerial robot visual navigation,

    S. Krishnan, B. Boroujerdian, W. Fu, A. Faust, and V . J. Reddi, “Air learning: a deep reinforcement learning gym for autonomous aerial robot visual navigation,” Machine Learning , vol. 110, pp. 2501–2540, 2021

    work page 2021

  28. [28]

    Autonomous navigation via deep reinforcement learning for resource constraint edge nodes using transfer learning,

    A. Anwar and A. Raychowdhury, “Autonomous navigation via deep reinforcement learning for resource constraint edge nodes using transfer learning,” IEEE Access , vol. 8, pp. 26 549–26 560, 2020

    work page 2020