Intelligent Multi-UAV Navigation in ITNTNs: A Hierarchical LLM Approach

Halim Yanikomeroglu; Hao Zhou; Hina Tabassum; Jianhua Pei; Ping Wang; Wael Jaafar; Zijiang Yan

REVIEW 2 major objections 5 minor 16 references

Intelligent Multi-UAV Navigation in ITNTNs: A Hierarchical LLM Approach

T0 review · 2 major / 5 minor · reviewed 2026-08-01 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read A two-tier LLM hierarchy cuts drone collisions and improves network throughput in 3D aerial highways.

desk verdict Plausible hierarchical LLM-DDQN architecture for multi-UAV ITNTN control, but the headline metrics (collision rate, throughput) are never measured; the evaluation is the weak link. read the letter →

arxiv 2607.18604 v1 pith:LN3HEWDQ submitted 2026-07-21 cs.RO cs.AIcs.LGcs.NIcs.SYeess.SY

Zijiang Yan , Hao Zhou , Wael Jaafar , Jianhua Pei , Ping Wang , Halim Yanikomeroglu , Hina Tabassum This is my paper · ORCID

classification cs.ROcs.AIcs.LGcs.NIcs.SYeess.SY

keywords multi-UAVnavigationaerialhighwayslargelanguagemodelsdeepreinforcementlearninghandovermanagementintegratedterrestrialandnon-terrestrialnetworkshierarchicalcontrolrewardshaping

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

This paper is trying to show that the complementary strengths of large language models and deep reinforcement learning can be combined in a hierarchy for controlling swarms of high-speed drones in 3D aerial highways. The key idea is to let a cloud-based language model handle slow, global decisions like balancing load across base stations, while smaller onboard language models convert each drone's local situation into a short-term tactical goal by adjusting the weights of a multi-objective reward. A fast reinforcement-learning controller then executes those goals at 20 Hz. If the simulations are right, this arrangement produces fewer collisions, longer survival times, and better overall network throughput than reinforcement learning alone, and it offers a template for injecting semantic reasoning into real-time control loops.

What carries the argument

The load-bearing mechanism is the cognitive reflection loop: an edge LLM converts a discretized semantic state (distance zones, link-stability labels) into a new scalarization of the multi-objective reward every 1 s, thereby steering the fast Deep Double Q-Network without injecting action commands directly. This turns the LLM into a reward shaper, and the HAPS cloud LLM adds a slower loop that issues load-balancing meta-actions to prevent base-station saturation.

What would settle it

A controlled comparison where the edge LLM's weight adjustments are replaced either by fixed weights or by a simple rule-based scheduler that maps the same semantic state (distance zones, link stability) to weights. If the rule-based version matches or beats the LLM version in collision rate and throughput, the paper's claim that LLM zero-shot reasoning drives the improvement would be falsified; if instead the LLM version clearly wins, the claim gains support.

Watch

Extended reading notes

Core claim

The paper sets out to show that the latency of large language models need not disqualify them from real-time drone control. It proposes a hierarchy: a cloud LLM on a high-altitude platform reasons every five seconds about global load balancing, while each drone runs a smaller edge LLM once per second that reads a semantic description of its local state and outputs a new weighting of the reward components (transit progress, link quality, collision penalty). A Deep Double Q-Network then acts on that weighted reward at 20 Hz, picking rotor speeds and network handovers. In simulation, this stack converges to higher total reward, survives longer without collisions, and sustains lower handover pro

Load-bearing premise

The framework assumes that the reward weights tuned by an onboard language model every second will not destabilize the deep reinforcement learning controller that must keep learning from them; no convergence or ablation evidence directly supports that.

Editorial extensions

If this is right

The LLM-guided framework overtakes the DRL baselines around episode 1,500 and reaches a higher steady-state reward.
Collisions that terminate episodes early in the baseline are largely avoided, so survival step counts approach the maximum episode length.
Handover probability stays lowest and flattest as the number of drones grows from 5 to 30.
The communication reward is highest once multiple base stations create interference, because the HAPS meta-controller coordinates associations.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

Inference: The reward-reshaping pattern could be reused in other multi-agent settings where a slow semantic layer must guide a fast control layer, e.g., connected ground vehicles or warehouse robots, without coupling the LLM to the control loop.
Inference: The paper does not ablate the edge-LLM's weight tuning against a fixed or rule-based scheduler; if a simple heuristic that maps semantic zones to weights performed equally, the observed gains would land on the reward adaptation, not on LLM reasoning. A direct comparison would settle this.
Inference: Because the edge LLM retrieves few-shot examples from local memory, the framework may exhibit self-improvement over time; a long-horizon deployment could make the weight adjustments more consistent, a property the current simulations do not measure.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Desk Editor's Note

read the letter

The architecture is a sensible engineering idea, and the paper does a good job motivating the timescale separation: edge-LLM latency at 0.6 s, cloud LLM at 2.55 s, and a 20 Hz control loop. Putting the big LLM on the HAPS for slow global load balancing and a small LLM on each UAV for tactical reward weighting is a concrete, non-obvious way to inject semantic reasoning into real-time control. The prompt template and the few-shot retrieval scheme are also useful details. This clearly extends the authors' prior hybrid LLM-DDQN work to the multi-UAV case with a HAPS meta-controller.

The problem is the evaluation. The abstract claims the architecture 'significantly reduces collision rates and improves aggregate system throughput,' but Figure 3 shows total reward, survival step count, communication reward, and handover probability. None of these is a direct collision-rate or throughput curve. Survival step count is a proxy that can be inflated by conservative flying, and the communication reward is a scalarized weighted sum, not Mbps. There are no error bars, confidence intervals, or multi-seed statistics, so 'significantly' is not supported in any statistical sense. This is the load-bearing gap: without direct metrics, the paper's strongest claim is unverified.

The reader's concern about non-stationary rewards is also real. The edge-LLM retunes alpha_1, alpha_2, alpha_3 every second, which makes the DDQN's reward a moving target. Standard convergence guarantees do not apply, and there is no ablation against fixed weights or a rule-based scheduler. Without that ablation, you cannot tell whether the gains come from LLM reasoning or just from dynamic reward shaping. The HAPS meta-controller's directives (Offload/Recall/Idle) are also described only abstractly; it is not clear how they are enforced locally.

That said, the core comparison is not circular. The baselines are external, and the architecture itself is coherent. The missing pieces are measurement and attribution, not conceptual integrity. I think the paper deserves a serious referee because the architecture is novel and the latency analysis is thoughtful, but it needs major revision: direct collision/throughput results, error bars, and an ablation against a simpler weight scheduler. Without those, the claims should be toned down. I would not cite the results yet, but I would point colleagues to the design ideas.

Referee Report

2 major / 5 minor

Summary. The paper proposes a hierarchical LLM-driven control framework for multi-UAV navigation in ITNTNs. A cloud-based LLM on a HAPS performs slow-timescale global load balancing, while lightweight edge-LLMs on UAVs periodically adjust the reward weights of a fast-timescale DDQN controller. The problem is formulated as a hierarchical multi-objective POMDP. Simulations in a gym-pybullet-drones environment compare the proposed framework against DDQN and Envelope MORL baselines. The abstract claims 'significantly reduces collision rates and improves aggregate system throughput,' but the reported results are proxy metrics only.

Significance. If the claimed gains are real, the architecture would demonstrate a practical way to inject strategic, semantic reasoning into real-time multi-UAV control without violating latency constraints. The timescale decomposition (HAPS at 5 s, edge-LLM at 1 s, DDQN at 0.05 s) is well motivated, and the use of a high-fidelity physics simulator is a strength. However, the current evidence does not support the headline empirical claims, and the non-stationary reward tuning raises unresolved stability concerns.

major comments (2)

[Abstract and Section V-B, Fig. 3] The abstract claims that the architecture 'significantly reduces collision rates and improves aggregate system throughput,' but neither metric is directly measured or reported. Fig. 3 shows total reward, survival step count, communication reward, and handover probability. Survival step count is only a proxy for collision avoidance and can increase under conservative flying that sacrifices throughput; communication reward is a weighted combination that includes the handover penalty, not throughput in Mbps. No error bars, confidence intervals, or multi-seed statistics are provided, so 'significantly' is not supported. The paper must report direct collision rates (e.g., per episode or per flight hour) and aggregate throughput (e.g., bps/Hz or Mbps), with multiple random seeds and statistical significance tests.
[Section IV-B, Eq. (7)] The edge-LLM dynamically adjusts the reward weights α_{1,2,3} every T_LLM = 1 s while the DDQN continues training on the resulting reward. This makes the reward non-stationary, and standard DDQN convergence guarantees no longer apply. The paper offers no convergence analysis and no ablation against fixed weights or a simple rule-based weight scheduler. Without this comparison, the gains in Fig. 3 could be attributed to reward shaping rather than to LLM reasoning. Please provide learning curves under fixed weights, a hand-crafted scheduling rule, and the proposed LLM-based tuning, with multiple seeds.

minor comments (5)

[Section V-A and Table I] The number of UAVs is reported inconsistently: the text says M ∈ {10,20,30}, while Table I lists M ∈ {5,10,15,20,25,30}. Clarify which set is used for the experiments in Fig. 3.
[Section V-A, references] The DDQN baseline is cited to [4] and [15], but [15] appears to be a paper on programmable data-planes, unrelated to deep Q-networks. This citation seems erroneous.
[Eq. (7) and surrounding text] There is a typo: 'rewrards' should be 'rewards'. Also, the notation R_m^t is overloaded with the communication rate R^{m,c}_t; consider renaming one to avoid confusion.
[Fig. 3 caption] The caption says 'Training convergence and scalability evaluation,' but panels (c) and (d) are not training-convergence curves. Adjust the caption to reflect that (a)-(b) are training curves and (c)-(d) are parameter sweeps.
[Section IV-D and abstract] The abstract and introduction emphasize 'zero-shot' reasoning, but the edge-LLM uses few-shot examples retrieved from a memory buffer. The term 'zero-shot' is therefore inaccurate; please qualify it (e.g., 'zero-shot strategic directives at the HAPS' vs. 'few-shot in-context tuning at the edge').

Circularity Check

0 steps flagged · score 0.0 of 10

No material circularity: the hierarchical LLM/DDQN architecture is evaluated against external baselines, and the reward-based metrics and self-citations do not reduce the central claim to its inputs.

full rationale

I examined the derivation chain for definitional reductions. The UAV reward in Eq. (7) is a weighted sum of transit reward, handover-aware data rate, and collision penalty; the edge-LLM adjusts the weights α and the DDQN maximizes that same scalarized reward. Fig. 3 plots components of that reward (transportation reward, communication reward) plus proxy metrics (survival step count, handover probability), while the abstract claims direct gains in collision rates and aggregate throughput. This is a reporting/measurement gap, not circularity: the plotted quantities are objective terms, but the comparison is against DDQN and Envelope-MORL baselines trained under the same reward structure, so the gains are not forced by construction. The few-shot memory retrieval in Section IV-D is self-referential insofar as 'successful' weight configurations come from the agent's own buffer, but no fitted parameter is renamed as a prediction; it is an algorithmic design choice, not a claimed first-principles output. Self-citations ([2], [4], [9], [11], [16]) are used for background, baselines, or prior architecture context; no load-bearing uniqueness theorem or ansatz is imported from prior work. Therefore, no circular step meeting the quoted-reduction standard is present.

Assumptions & free parameters 5 free parameters · 4 assumptions · 0 invented entities

The paper introduces no new physical entities; the LLM controllers are system components. The central claim rests heavily on hand-chosen reward weights, latency estimates, and the unverified assumption that reward re-weighting preserves DDQN learnability.

free parameters (5)

Initial UAV reward weights α_{1,2,3} = {1.0, 0.1, 0.2}
Hand-chosen in Table I; these scale transit, network, and safety terms in Eq. (7). The edge-LLM then adjusts them, so the baseline behavior depends on this choice.
HAPS reward weights η_{1,2} = {1.0, 50.0}
Hand-chosen in Table I; the large 50.0 penalty on congestion drives load-balancing behavior in Eq. (8).
Handover penalty coefficient γ = 5.0
Chosen in Table I and used in Eq. (3); it defines when a link is 'Hand-over' (WR < 0) and affects the edge-LLM's state descriptors.
Safety reflection threshold ρ_thresh = -10.0
Threshold in Section IV-B that triggers edge-LLM cognitive reflection; no sensitivity analysis.
LLM reasoning interval T_LLM and HAPS interval T_HAPS = 1.0s, 5.0s
Chosen based on estimated latencies t_edge=0.6s, t_cloud=2.55s; the validity of these latency estimates is assumed.

assumptions (4)

domain assumption The kinematic model in Eqs. (4)-(5) accurately represents UAV flight for control purposes
Adopted from gym-pybullet-drones [14]; the paper does not validate the model against real flight data, and collision safety depends on it.
ad hoc to paper LLM inference can be treated as a fixed-latency black box (t_edge=0.6s, t_cloud=2.55s) with reliable output
Section V-A states these values; the paper assumes the Qwen3.5-9B/122B models produce semantically correct reward weights and meta-actions, with no error analysis or guarantee.
ad hoc to paper The scalarized reward in Eq. (7) with time-varying LLM-set weights is a valid objective for DDQN training
The DDQN learns under non-stationary reward weights updated every 1s; standard convergence proofs do not apply, and the paper provides no analysis.
domain assumption SINR/rate model Eqs. (1)-(3) capture the essential ITNTN handover dynamics
Standard models from prior literature; parameters are simulated, not measured.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Intelligent Multi-UAV Navigation in ITNTNs: A Hierarchical LLM Approach." pith.science (2026). https://pith.science/paper/LN3HEWDQ

@misc{pith2026260718604,
  author       = {Pith},
  title        = {Pith review of: Intelligent Multi-UAV Navigation in ITNTNs: A Hierarchical LLM Approach},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/LN3HEWDQ}},
  note         = {Machine review of arXiv:2607.18604}
}

read the original abstract

The deployment of high-speed Uncrewed Aerial Vehicles (UAVs) in 3D aerial highways necessitates robust coordination of physical flight kinematics and multi-tier network handovers. While Deep Reinforcement Learning (DRL) offers rapid tactical control, it lacks the zero-shot strategic reasoning required to quickly adapt to dynamic Integrated Terrestrial and Non-Terrestrial Networks (ITNTNs). Conversely, Large Language Models (LLMs) excel at semantic reasoning but suffer from high inference latency, rendering them unsuitable for real-time aerodynamic control. To bridge this gap, we propose a novel Hierarchical LLM-driven control framework. A massive cloud-based LLM deployed on a High-Altitude Platform Station (HAPS) manages slow-timescale global load balancing, while lightweight edge-LLMs on individual UAVs translate local observations into tactical sub-goals. These sub-goals guide a fast-timescale physical DRL controller to execute collision-free, handover-aware trajectories. Simulation results demonstrate that our agentic architecture significantly reduces collision rates and improves aggregate system throughput compared to existing baselines.

Figures

Figures reproduced from arXiv: 2607.18604 by the authors.

**Figure 1.** 3D aerial network model. problem as a H-MO-POMDP. The proposed dual-timescale hierarchical LLM framework and its cognitive reasoning modules are introduced in Section IV. Section V presents the numerical results and performance evaluations. Finally, Section VI concludes the paper. II. SYSTEM MODEL A. 3D Aerial Highway and ITNTN Architecture As depicted in [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗

**Figure 2.** Hierarchical control for UAVs in an ITNTN. [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗

**Figure 3.** Training convergence and scalability evaluation of the proposed Cloud-Edge Agentic framework. [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

16 extracted references · 2 linked inside Pith

[1]

3D aerial highway: The key enabler of the retail industry transformation,

N. Cherifet al., “3D aerial highway: The key enabler of the retail industry transformation,”IEEE Commun. Mag., vol. 59, no. 9, pp. 65– 71, 2021

2021
[2]

Multi-UA V speed control with collision avoidance and handover-aware cell association: DRL with action branching,

Z. Yanet al., “Multi-UA V speed control with collision avoidance and handover-aware cell association: DRL with action branching,” inProc. IEEE Global Commun. Conf. (GLOBECOM), 2023, pp. 5067–5072

2023
[3]

HAPS-ITS: Enabling future its services in trans-continental highways,

W. Jaafar and H. Yanikomeroglu, “HAPS-ITS: Enabling future its services in trans-continental highways,”IEEE Commun. Mag., vol. 60, no. 10, pp. 80–86, 2022

2022
[4]

Reinforcement learning for joint V2I network selection and autonomous driving policies,

Z. Yan and H. Tabassum, “Reinforcement learning for joint V2I network selection and autonomous driving policies,” inProc. IEEE Global Commun. Conf. (GLOBECOM), 2022, pp. 1241–1246

2022
[5]

RL-based cargo-UA V trajectory planning and cell association for minimum handoffs, disconnectivity, and energy con- sumption,

N. Cherifet al., “RL-based cargo-UA V trajectory planning and cell association for minimum handoffs, disconnectivity, and energy con- sumption,”IEEE Trans. Veh. Technol., vol. 73, no. 5, pp. 7304–7309, 2024

2024
[6]

Dynamics of laser-charged UA Vs: A battery perspective,

W. Jaafar and H. Yanikomeroglu, “Dynamics of laser-charged UA Vs: A battery perspective,”IEEE Internet Things J., vol. 8, no. 13, pp. 10 573–10 582, 2020

2020
[7]

CVaR-based variational quantum optimization for user association in handoff-aware vehicular networks,

Z. Yanet al., “CVaR-based variational quantum optimization for user association in handoff-aware vehicular networks,” inProc. IEEE Int. Conf. Commun. (ICC), 2025, pp. 6088–6093

2025
[8]

Aerial Agentic AI: Synergizing LLM and SLM for Low-Altitude Wireless Networks,

L. Donget al., “Aerial Agentic AI: Synergizing LLM and SLM for Low-Altitude Wireless Networks,”arXiv preprint arXiv:2603.22866, 2026

arXiv 2026

Show all 16 references

[9]

Hierarchical and Collaborative LLM-Based Control for Multi-UA V Motion and Communication in Integrated Terrestrial and Non-Terrestrial Networks,

Z. Yanet al., “Hierarchical and Collaborative LLM-Based Control for Multi-UA V Motion and Communication in Integrated Terrestrial and Non-Terrestrial Networks,”arXiv preprint arXiv:2506.06532, 2025

2025 arXiv
[10]

BDFL: A byzantine-fault-tolerance decentralized federated learning method for autonomous vehicle,

J.-H. Chenet al., “BDFL: A byzantine-fault-tolerance decentralized federated learning method for autonomous vehicle,”IEEE Trans. Veh. Technol., vol. 70, no. 9, pp. 8639–8652, 2021

2021
[11]

Hybrid LLM-DDQN-Based Joint Optimization of V2I Communication and Autonomous Driving,

Z. Yanet al., “Hybrid LLM-DDQN-Based Joint Optimization of V2I Communication and Autonomous Driving,”IEEE Wireless Commun. Lett., vol. 14, no. 4, pp. 1214–1218, 2025

2025
[12]

Qwen3 technical report,

A. Yanget al., “Qwen3 technical report,”arXiv preprint arXiv:2505.09388, 2025

2025 arXiv
[13]

Cellular-connected UA V in next-generation wireless net- works,

N. Cherif, “Cellular-connected UA V in next-generation wireless net- works,” Ph.D. dissertation, Univ. Ottawa, 2022

2022
[14]

Learning to fly—a gym environment with pybullet physics for reinforcement learning of multi-agent quadcopter control,

J. Paneratiet al., “Learning to fly—a gym environment with pybullet physics for reinforcement learning of multi-agent quadcopter control,” in Proc. IEEE/RSJ Int. Conf. Intell. Robots Syst. (IROS), 2021, pp. 7512– 7519

2021
[15]

A machine learning-based toolbox for p4 pro- grammable data-planes,

K. Zhanget al., “A machine learning-based toolbox for p4 pro- grammable data-planes,”IEEE Trans. Netw. Serv. Manag., vol. 21, no. 4, pp. 4450–4465, 2024

2024
[16]

Generalized multi-objective reinforcement learning with envelope updates in URLLC-enabled vehicular networks,

Z. Yan and H. Tabassum, “Generalized multi-objective reinforcement learning with envelope updates in URLLC-enabled vehicular networks,” IEEE Trans. Veh. Technol., vol. 74, no. 11, pp. 17 666–17 682, 2025

2025

Pith tools

Reviewed August 1, 2026 · model on record in the stance chip above.

[1] [1]

3D aerial highway: The key enabler of the retail industry transformation,

N. Cherifet al., “3D aerial highway: The key enabler of the retail industry transformation,”IEEE Commun. Mag., vol. 59, no. 9, pp. 65– 71, 2021

2021

[2] [2]

Multi-UA V speed control with collision avoidance and handover-aware cell association: DRL with action branching,

Z. Yanet al., “Multi-UA V speed control with collision avoidance and handover-aware cell association: DRL with action branching,” inProc. IEEE Global Commun. Conf. (GLOBECOM), 2023, pp. 5067–5072

2023

[3] [3]

HAPS-ITS: Enabling future its services in trans-continental highways,

W. Jaafar and H. Yanikomeroglu, “HAPS-ITS: Enabling future its services in trans-continental highways,”IEEE Commun. Mag., vol. 60, no. 10, pp. 80–86, 2022

2022

[4] [4]

Reinforcement learning for joint V2I network selection and autonomous driving policies,

Z. Yan and H. Tabassum, “Reinforcement learning for joint V2I network selection and autonomous driving policies,” inProc. IEEE Global Commun. Conf. (GLOBECOM), 2022, pp. 1241–1246

2022

[5] [5]

RL-based cargo-UA V trajectory planning and cell association for minimum handoffs, disconnectivity, and energy con- sumption,

N. Cherifet al., “RL-based cargo-UA V trajectory planning and cell association for minimum handoffs, disconnectivity, and energy con- sumption,”IEEE Trans. Veh. Technol., vol. 73, no. 5, pp. 7304–7309, 2024

2024

[6] [6]

Dynamics of laser-charged UA Vs: A battery perspective,

W. Jaafar and H. Yanikomeroglu, “Dynamics of laser-charged UA Vs: A battery perspective,”IEEE Internet Things J., vol. 8, no. 13, pp. 10 573–10 582, 2020

2020

[7] [7]

CVaR-based variational quantum optimization for user association in handoff-aware vehicular networks,

Z. Yanet al., “CVaR-based variational quantum optimization for user association in handoff-aware vehicular networks,” inProc. IEEE Int. Conf. Commun. (ICC), 2025, pp. 6088–6093

2025

[8] [8]

Aerial Agentic AI: Synergizing LLM and SLM for Low-Altitude Wireless Networks,

L. Donget al., “Aerial Agentic AI: Synergizing LLM and SLM for Low-Altitude Wireless Networks,”arXiv preprint arXiv:2603.22866, 2026

arXiv 2026