TRUAV: Distributed Multi-Agent Reinforcement Learning for Trajectory Planning and Routing Enhancement in UAV-Aided IoT-Enabled VANETs

Chang Liu; Lin Zhang; Muhammad Umar Farooq Qaisar; Shehzad Ashraf Chaudhry; Wajdy Othman; Zhen Chen

REVIEW 4 major objections 7 minor 14 references

TRUAV: Distributed Multi-Agent Reinforcement Learning for Trajectory Planning and Routing Enhancement in UAV-Aided IoT-Enabled VANETs

T0 review · 4 major / 7 minor · reviewed 2026-07-30 · grok-4.5

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

Pith's one-line read Independent local Q-learners can plan UAV paths for VANET routing as well as a centralized deep critic, with less delay and energy.

desk verdict Solid distributed-UAV systems paper with a real sim comparison, but the headline wins over centralized MASAC rest on an unverifiable baseline and energy/delay attributions the model never actually charges. read the letter →

arxiv 2607.23734 v1 pith:PTZFUHBO submitted 2026-07-26 cs.NI cs.LG

Muhammad Umar Farooq Qaisar , Lin Zhang , Zhen Chen , Wajdy Othman , Shehzad Ashraf Chaudhry , Chang Liu This is my paper · ORCID

classification cs.NIcs.LG

keywords UAVcommunicationsVANETInternetofThingsdistributedreinforcementlearningmulti-agentsystemstrajectoryoptimizationpotential-game-inspiredrewardaerialrelaying

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

Centralized deep reinforcement learning can jointly plan UAV trajectories and VANET routing, but it needs continuous global state from every agent, which does not scale under urban bandwidth and energy limits. TRUAV instead gives each UAV its own tabular Q-learning agent that sees only local vehicle density, packet queues, neighbor UAV positions, and distance to the nearest roadside unit. A hand-weighted reward that scores successful relays and useful coverage while penalizing energy use and spatial overlap is meant to align those selfish updates with network-wide routing utility, in the spirit of a potential game, without any shared critic or coordination messages. In a 5 km by 5 km simulation with 200 vehicles and five UAVs, the distributed agents match the centralized baseline on coverage and packet delivery while cutting average relay delay and episode energy. The practical claim is that routing-aware aerial IoT relays can be learned onboard from beacons alone.

What carries the argument

The local reward in equation (1): a weighted sum of relay success, coverage, energy cost and an overlap penalty, designed so that asynchronous independent Q-updates push the swarm toward spatially diverse, routing-useful positions without global state exchange.

What would settle it

Rerun the 5000 m urban scenario with the same five UAVs and 200 vehicles; if independent tabular Q-learning under the stated reward falls more than a few points behind the centralized soft actor-critic baseline on coverage or packet delivery, or fails to reduce delay and energy, the central claim does not hold.

Watch

Extended reading notes

Core claim

TRUAV shows that independent tabular Q-learning agents, each using only locally observable vehicle density, queue backlog, neighbor UAV positions and RSU distance, plus a potential-game-inspired reward that rewards relays and coverage while penalizing energy and overlap, can achieve network coverage and packet delivery ratios comparable to a centralized multi-agent soft actor-critic baseline while improving relay delay and energy consumption.

Load-bearing premise

That the hand-tuned local reward weights and overlap penalty are aligned enough with global routing utility that independent Q-updates will steer the swarm near network-level optima without a formal potential-function proof or shared critic.

Editorial extensions

If this is right

UAV trajectory control for VANET/IoT relaying need not broadcast global state or run a central critic at training or runtime.
Lightweight onboard Q-tables on discretized local features become a practical alternative to deep multi-agent critics for small UAV swarms.
An overlap penalty in each agent’s reward can spread aerial relays without explicit coordination messages.
Coupling the reward to successful packet relays rather than pure geometric coverage ties flight paths to actual traffic demand.
The same local-observation pattern can be extended to heterogeneous UAV fleets and predictive traffic models as sketched in the paper’s future work.

Reading between the lines

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

If local tabular agents already match centralized DRL on this scale, the next stress test is whether the same reward still coordinates when swarm size or vehicle density grows enough that neighbor beacons become sparse or contested.
The discretized state tuple is what makes onboard tabular learning feasible; any real deployment will hinge on whether those bins stay informative under realistic sensing noise and beacon loss.
Federated distillation of Q-tables, listed as future work, would be a natural next step if privacy or backhaul limits bar even lightweight policy sharing.

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 useful takeaway is narrow but real: independent tabular Q-learning with a local state (density, queues, neighbor positions, RSU distance) and a composite reward can, in their 5 km urban sim with 200 vehicles and 5 UAVs, match a centralized joint trajectory-routing DRL baseline on coverage and PDR while looking better on delay and energy. That is honest engineering progress on the control-plane side of aerial IoT relays, not a new theory class.

What they do well is keep the design deployable. Local observations only, no global critic, compact discrete state for onboard tabular updates, and an overlap penalty that spreads the swarm without coordination messages. The system model (CFM mobility, probabilistic LoS, hover/move power) is standard and readable. Related work is mostly fair about Chen et al. and the static/dynamic/learning phases. Metrics are external (coverage, PDR, delay, energy), so circularity is low.

Soft spots, in proportion. The potential-game story is mostly branding: Eq. (1) is hand-tuned (α=1, β=0.25, γ=0.15, δ=0.20 via “manual tuning and a brief sensitivity sweep”) with no potential-function proof, only a citation to unrelated MARL work. State binning is underspecified. No code, no error bars despite 10 seeds; the reported margins (~1 pt coverage, ~3 ms delay, ~1 kJ) could be seed noise. The stress-test lands on the comparison: delay/energy wins are attributed to “no controller latency / no global state broadcasts,” yet §III-C’s energy model only counts propulsion plus packet TX/RX, and nothing models aggregation latency. Either the baseline was asymmetrically handicapped or the mechanism is not simulated. Their own Python re-implementation of MASAC with undocumented tuning budget makes a fully-observed centralized method losing on every metric hard to trust as written. Duplicate refs [8]/[12] are minor sloppiness.

Who it is for: people building multi-UAV VANET/IoT control who care about distributed training under bandwidth limits. Not for theory or anyone needing reproducible numbers tomorrow.

I would send it to peer review. A serious referee can demand ablations, CIs, baseline tuning parity, and either model the claimed overheads or drop the attribution. Worth engaging if you work this area; cite only after those fixes or after you re-run it yourself.

Referee Report

4 major / 7 minor

Summary. The manuscript presents TRUAV, a distributed multi-agent reinforcement learning framework for joint UAV trajectory planning and routing in UAV-aided VANETs. Each of 5 UAVs runs an independent tabular Q-learning agent on a local state (vehicle density, queue length, neighbor UAV positions, RSU distance), with a hand-weighted reward (Eq. 1) combining relay success, coverage, energy, and an overlap penalty described as "potential-game-inspired." In a custom Python simulator (5000 m × 5000 m, 200 vehicles, 4 RSUs, IEEE 802.11p), the authors report that TRUAV matches a centralized multi-agent SAC baseline [12] on coverage (~94%) and PDR (~91.5%) while achieving lower relay delay (~25.2 ms vs ~28.1 ms at 25 m/s) and ~19% lower episode energy, and outperforms a trajectory-knowledge routing baseline [10] on all metrics. Results are means over 10 seeds with 500-episode smoothing.

Significance. If the results hold, the paper makes a useful practical point for the UAV-IoT/VANET community: lightweight tabular Q-learning with purely local observations can be competitive with centralized deep RL on coverage and delivery ratio, which matters for deployments where global state aggregation is infeasible. The joint treatment of trajectory and routing demand (rewarding actual packet relays rather than geometric coverage alone) is a sensible design contribution, and the simulation study spans a reasonably large scenario (200 vehicles, 5×5 km). However, the headline claims are comparative and rest entirely on the authors' own simulator and their own re-implementation of the baselines, with no released code, no reported variance, and — critically — an energy/delay advantage attributed to mechanisms (controller latency, global state broadcasts) that do not appear in the stated system model. The significance is therefore conditional on the authors substantiating the comparison.

major comments (4)

[§VI-B, §VI-C vs. §III-C] The delay and energy advantages are attributed to mechanisms that are not in the model. §VI-B attributes TRUAV's delay win to 'immediate local action selection without controller latency or global state aggregation,' and §VI-C attributes the ~19% energy win to 'eliminating global state broadcasts.' Yet the energy model in §III-C accounts only for hover/cruise propulsion power plus per-packet TX/RX communication energy, and nothing in §III or §V defines a controller-aggregation latency term or a state-broadcast energy cost. Either these penalties were charged to the centralized baseline in the accounting without being modeled (an asymmetric handicap), or the claimed mechanism was never simulated and the attribution is speculative. The authors must either (a) extend §III-C with an explicit, parameterized model of control-plane latency and broadcast energy and show the contribution of each
[§V, Table I; Fig. 3a–d] No uncertainty quantification. §V states results are means over 10 seeds, but no standard deviations, confidence intervals, or error bars appear in Fig. 3 or the text. The central comparative gaps are small: ~1 percentage point of coverage at 15 m/s and ~2.9 ms of delay at 25 m/s. With 10 seeds and 500-episode smoothing, it is entirely plausible these gaps are within seed variance. The claims 'comparable coverage/PDR' and 'improved delay/energy' cannot be evaluated without per-point variance or a significance test; error bars on Fig. 3a–d are the minimum requirement.
[§V (baselines), §VI] Baseline provenance and tuning symmetry are undocumented. The MASAC baseline is the authors' own Python re-implementation of [12]; no code is released, no hyperparameter tuning budget for the baseline is stated, and its network architectures, training steps, and reward are unspecified. Meanwhile TRUAV's reward weights (α=1.0, β=0.25, γ=0.15, δ=0.20) were set by 'manual tuning and a brief sensitivity sweep' (§IV-C) — the sweep itself is not reported. A fully-observed centralized deep RL method underperforming a local-observation tabular method on every metric is surprising; the paper must document that the baseline was competently tuned (hyperparameters, training budget, sensitivity of the baseline's results) or the comparative claim is not evaluable. Releasing the simulator and baseline code would resolve most of this.
[§IV-E, Eq. (1)] The potential-game justification is asserted, not shown. §IV-E claims the reward in Eq. (1) is 'closely aligned' with a global potential so that independent Q-updates approach network-level optima, citing [14] for convergence. But [14] is an evolutionary role-learning MARL paper, not a potential-game convergence result, and no potential function is constructed or verified for Eq. (1) (e.g., the standard requirement that each agent's utility difference equal the potential difference under unilateral deviation). Since the overlap penalty and relay reward are agent-local, alignment is not automatic. Either provide a potential-function argument (even under simplifying assumptions) or retitle the design 'potential-game-inspired heuristic' and drop the convergence claim, which is load-bearing for the framing that independent learners coordinate without messages.

minor comments (7)

[References] References [8] and [12] are identical (same authors, title, journal, volume, pages, year). One should be removed and citations consolidated.
[§IV-A] The state discretization is underspecified: the number of grid cells, the ordinal levels for density/queue/neighbor distance, and the resulting Q-table size are not given. These determine whether 5000 episodes suffice for tabular convergence and should be in Table I.
[Fig. 3] Figure 3 subplots have no visible axis labels/units in the provided text and no legend description beyond the caption; please ensure axes (UAV velocity in m/s; coverage/PDR in %; delay in ms; energy in kJ) and method legends are legible, and add error bars per the major comment.
[§V] The packet traffic model (generation rate, packet size, destination distribution, queueing discipline) that drives PDR and delay is never specified; these parameters materially affect all four metrics and belong in Table I.
[§I, ¶6] The overhead arithmetic ('six UAVs associated with 50-dimensional state vector at 10 Hz ... 15000 values a second') should be stated precisely (6×50×10 = 3000 values/s; 15000 presumably includes something else — clarify).
[§IV-B] Typos/grammar: 'to choose an actions: at the start of training it selects random action'; 'converges to a stable configurations' (§IV-E); 'discretized states representation' (§VII). A proofreading pass is needed.
[§III-B] The LoS probability model and path-loss exponents are described only qualitatively ('depends on the elevation angle', 'additional term for urban clutter'); give the actual equations or cite the specific model used so the link-budget is reproducible.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: empirical MARL systems paper with external benchmarks; reward design is intentional engineering, not a self-defining derivation.

full rationale

TRUAV is an empirical multi-agent RL systems paper. Its load-bearing claim is comparative simulation performance (coverage, PDR, relay delay, energy) of independent tabular Q-learning against a centralized MASAC baseline and a trajectory-knowledge baseline over a stated urban scenario. Those metrics are standard external network benchmarks, not quantities algebraically forced by a fitted parameter. The local reward in Eq. (1) deliberately includes relay, coverage, energy, and overlap terms with hand-tuned weights; measuring improvement on related KPIs is ordinary RL evaluation, not a self-definitional ‘prediction.’ The ‘potential-game-inspired’ narrative in §IV-E asserts alignment of local rewards with global routing utility and cites Tu et al. [14] for decentralized Q-learning stability—[14] is not by the present authors and is not used as a uniqueness theorem that forbids alternatives—so this is weak justification at worst, not a self-citation chain. Authors’ own prior citations [1],[2] address trust/ISAC topics and are not load-bearing for the trajectory/routing results. No fitted input is renamed a first-principles prediction; no ansatz is smuggled in via overlapping-author uniqueness; no known empirical law is merely relabeled. Per the analyzer default, this is an honest non-finding.

Assumptions & free parameters 5 free parameters · 5 assumptions · 2 invented entities

The central performance claim depends on simulation modeling choices, a hand-weighted reward treated as aligned with global routing utility, and the sufficiency of heavily discretized local observations for joint trajectory-routing quality. Little is derived; most load-bearing content is parameterization and domain modeling rather than proved coordination guarantees.

free parameters (5)

Reward weights α, β, γ, δ = α=1.0, β=0.25, γ=0.15, δ=0.20
Explicitly set to 1.0, 0.25, 0.15, 0.20 by manual tuning and a brief sensitivity sweep; they define the tradeoff the agents optimize and thus shape all reported metrics.
Q-learning hyperparameters (lr, γ_disc, ε schedule) = lr=0.1, discount=0.95, ε:1.0→0.05
Learning rate 0.1, discount 0.95, ε from 1.0 to 0.05 over 5000 episodes control convergence and final policy quality in all runs.
State discretization bins (density, queue, neighbor distance, grid)
Tabular Q-learning requires ordinal/grid encoding; bin edges are not fully specified but determine the entire representable policy class.
Action set cardinality and speed levels = 16 actions; 5–25 m/s
16 actions (8 directions × 2 speeds) and speed range 5–25 m/s are design choices that bound agility vs learning complexity.
UAV power model (hover/move) = 200 W hover, 250 W move
Hover 200 W and move 250 W directly drive the energy metric where TRUAV claims ~19% improvement.

assumptions (5)

domain assumption Local observations (vehicle density, queue backlog, neighbor UAV positions, distance to nearest RSU) are sufficient to learn trajectories that nearly match globally trained joint policies on network-wide routing utility.
Core premise of §IV-A and the distributed claim in the abstract; never proved, only supported by one simulation regime.
ad hoc to paper The shaped reward (1) is aligned enough with a global potential that independent Q-learning converges to near-optimal swarm configurations; overlap term acts as implicit repulsion without messaging.
§IV-E asserts potential-game inspiration and cites [14] for decentralized Q-learning stability; no potential function is exhibited for this reward.
domain assumption Urban vehicle motion is adequately represented by a car-following model on a regular two-lane grid with probabilistic turns.
§III-A; standard but load-bearing for claimed coverage tracking of “clusters.”
domain assumption Air–ground links follow probabilistic LoS plus urban clutter path loss and Shannon rates on IEEE 802.11p-like parameters; fixed 100 m altitude, 500 m range.
§III-B and Table I; determines coverage and PDR definitions used in claims.
domain assumption Asynchronous independent tabular updates remain stable under non-stationarity induced by other learning UAVs and mobile traffic.
§IV-D; classical independent Q-learning non-stationarity issue is acknowledged only lightly.

invented entities (2)

TRUAV reward (relay + coverage − energy − overlap)
purpose: Couple local trajectory choices to routing utility and spatial diversity without a central critic.
Composite reward is the paper’s main design object; success metrics partially overlap the reward terms, so independent evidence is only the held-out comparison to baselines under the same sim.
TRUAV local state tuple for tabular UAV agents
purpose: Make onboard tabular Q-learning feasible without neural nets or global state.
Specific feature bundle and discretization are proposed here; no external validation set or hardware trace.

how reviews work

0 comments

Cite this review

Pith. "Pith review of TRUAV: Distributed Multi-Agent Reinforcement Learning for Trajectory Planning and Routing Enhancement in UAV-Aided IoT-Enabled VANETs." pith.science (2026). https://pith.science/paper/PTZFUHBO

@misc{pith2026260723734,
  author       = {Pith},
  title        = {Pith review of: TRUAV: Distributed Multi-Agent Reinforcement Learning for Trajectory Planning and Routing Enhancement in UAV-Aided IoT-Enabled VANETs},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/PTZFUHBO}},
  note         = {Machine review of arXiv:2607.23734}
}

read the original abstract

Unmanned aerial vehicles (UAVs) have emerged as a key enabler of next-generation Internet of Things (IoT) ecosystems, offering flexible aerial relaying to extend connectivity across dynamic vehicular ad hoc networks (VANETs) in smart city environments. However, conventional centralized approaches for UAV trajectory planning require continuous global network state aggregation, making them impractical under bandwidth and energy constraints typical of dense urban deployments. In this article, we present TRUAV, a distributed multi-agent reinforcement learning framework based on independent tabular Q-learning for joint UAV trajectory planning and routing enhancement in UAV-aided VANETs. Each UAV is equipped with a local Q-learning agent that operates purely on locally observable information, including vehicle density, packet queue states, and neighbor UAV positions, thereby eliminating the need for global state exchange. A potential-game-inspired reward design encourages spatial diversity and routing-aware UAV positioning among interacting agents while accounting for energy consumption. Numerical simulations over a large urban area with 200 mobile vehicles show that the proposed TRUAV framework achieves network coverage and packet delivery ratios comparable to centralized deep reinforcement learning methods, while also improving relay delay and energy efficiency. Finally, we discuss emerging challenges and future research directions for distributed multi-agent UAV-assisted IoT systems.

Figures

Figures reproduced from arXiv: 2607.23734 by the authors.

**Figure 2.** Logical workflow of the proposed TRUAV framework [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗

**Figure 3.** Performance versus UAV velocity TABLE I: Simulation Parameters Parameter Value Description Environment Simulation area 5000 by 5000 m Urban coverage region Number of vehicles 200 Ground vehicles (CFM mobility) Number of UAVs 5 Aerial relay platforms Number of RSUs 4 Corner-mounted backhaul nodes UAV altitude 100 m Fixed flight level UAV Movement Speed range 5–25 m/s UAV horizontal velocity Communication range 500 m … view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

14 extracted references

[12]

Enhancing routing performance through trajectory planning with drl in uav-aided vanets,

J. Chen, D. Huang, Y . Wang, Z. Yu, Z. Zhao, X. Cao, Y . Liu, T. Q. S. Quek, and D. O. Wu, “Enhancing routing performance through trajectory planning with drl in uav-aided vanets,”IEEE Transactions on Machine Learning in Communications and Networking, vol. 3, pp. 517–533, 2025

2025
[10]

Opti- mized routing protocol through exploitation of trajectory knowledge for uav swarms,

Z. Zhou, J. Tang, W. Feng, N. Zhao, Z. Yang, and kai-kit Wong, “Opti- mized routing protocol through exploitation of trajectory knowledge for uav swarms,”IEEE Transactions on Vehicular Technology, vol. 73, pp. 15 499–15 512, 2024

2024
[14]

Adaptive role learning with evolutionary multi-agent re- inforcement learning for uav-vehicle collaboration in sparse mobile crowdsensing,

C. Tu, Z. Yu, J. Huang, F. Huang, Y . Wu, L. Han, L. Wang, and R. Huang, “Adaptive role learning with evolutionary multi-agent re- inforcement learning for uav-vehicle collaboration in sparse mobile crowdsensing,”IEEE Internet of Things Journal, 2025

2025
[1]

A robust trust management system for v2x networks integrating isac with blockchain smart contracts,

M. U. F. Qaisar, W. Yuan, L. Zhang, S. A. Chaudhry, G. Han, and Y . Zhang, “A robust trust management system for v2x networks integrating isac with blockchain smart contracts,”IEEE Transactions on Network and Service Management, vol. 23, pp. 2428–2439, 2026

2026
[2]

The role of isac in 6g networks: Enabling next-generation wireless systems,

M. U. F. Qaisar, W. Yuan, O. G ¨unl¨u, T. Riihonen, Y . Cui, L. Zhang, N. Gonz ´alez-Prelcic, M. di Renzo, and Z. Han, “The role of isac in 6g networks: Enabling next-generation wireless systems,”IEEE Transactions on Network Science and Engineering, vol. 13, pp. 7825– 7861, 2025

2025
[3]

Low-altitude wireless networks: A comprehensive survey,

W. Jun, Y . Yaoqi, Y . Weijie, L. Wenchao, W. Jiacheng, M. Tianqi, Z. Lin, C. Yuanhao, L. Fan, S. Genget al., “Low-altitude wireless networks: A comprehensive survey,”China Communications, vol. 23, no. 3, pp. 99–141, 2026

2026
[4]

Mobile edge computing via a uav- mounted cloudlet: Optimization of bit allocation and path planning,

S. Jeong, O. Simeone, and J. Kang, “Mobile edge computing via a uav- mounted cloudlet: Optimization of bit allocation and path planning,” IEEE Transactions on Vehicular Technology, vol. 67, no. 3, pp. 2049– 2063, 2017
[5]

Caching on the sky: A multiagent federated reinforcement learning approach for uav-assisted edge caching,

X. Li, J. Liu, X. Chen, J. Wang, and M. Pan, “Caching on the sky: A multiagent federated reinforcement learning approach for uav-assisted edge caching,”IEEE Internet of Things Journal, vol. 11, no. 17, pp. 28 213–28 226, 2024

2024

Show all 14 references

[6]

Deep reinforcement learning- based adaptive collision avoidance method for uav in joint operational airspace,

Y . Shen, X. Zhang, Y . Li, and W. Zhang, “Deep reinforcement learning- based adaptive collision avoidance method for uav in joint operational airspace,”Defence Technology, 2025

2025
[7]

Decoupled association with rate splitting multiple access in uav-assisted cellular networks using multi-agent deep reinforcement learning,

J. Ji, L. Cai, K. Zhu, and D. Niyato, “Decoupled association with rate splitting multiple access in uav-assisted cellular networks using multi-agent deep reinforcement learning,”IEEE Transactions on Mobile Computing, vol. 23, no. 3, pp. 2186–2201, 2023

2023
[9]

Predictive path planning of multiple uavs for effective network hotspot coverage,

J. Cho, S. Ki, and H. Lee, “Predictive path planning of multiple uavs for effective network hotspot coverage,”IEEE Transactions on Vehicular Technology, vol. 72, no. 12, pp. 16 683–16 700, 2023

2023
[11]

Joint uav placement and dependent task offloading in multi-uav mec networks: A graph attention enhanced drl approach,

C. Zhan, W. Liu, K. Song, R. Fan, J. Liu, and H. Hu, “Joint uav placement and dependent task offloading in multi-uav mec networks: A graph attention enhanced drl approach,”IEEE Transactions on Mobile Computing, vol. 25, pp. 5285–5301, 2026

2026
[13]

Graph soft actor–critic reinforcement learning for large-scale distributed multirobot coordination,

Y . Hu, J. Fu, and G. Wen, “Graph soft actor–critic reinforcement learning for large-scale distributed multirobot coordination,”IEEE transactions on neural networks and learning systems, vol. 36, no. 1, pp. 665–676, 2023

2023
[15]

Multi-agent rein- forcement learning based uav swarm communications against jamming,

Z. Lv, L. Xiao, Y . Du, G. Niu, C. Xing, and W. Xu, “Multi-agent rein- forcement learning based uav swarm communications against jamming,” IEEE Transactions on Wireless Communications, vol. 22, pp. 9063– 9075, 2023

2023

Pith tools

Reviewed July 30, 2026 · model on record in the stance chip above.

[1] [12]

Enhancing routing performance through trajectory planning with drl in uav-aided vanets,

J. Chen, D. Huang, Y . Wang, Z. Yu, Z. Zhao, X. Cao, Y . Liu, T. Q. S. Quek, and D. O. Wu, “Enhancing routing performance through trajectory planning with drl in uav-aided vanets,”IEEE Transactions on Machine Learning in Communications and Networking, vol. 3, pp. 517–533, 2025

2025

[2] [10]

Opti- mized routing protocol through exploitation of trajectory knowledge for uav swarms,

Z. Zhou, J. Tang, W. Feng, N. Zhao, Z. Yang, and kai-kit Wong, “Opti- mized routing protocol through exploitation of trajectory knowledge for uav swarms,”IEEE Transactions on Vehicular Technology, vol. 73, pp. 15 499–15 512, 2024

2024

[3] [14]

Adaptive role learning with evolutionary multi-agent re- inforcement learning for uav-vehicle collaboration in sparse mobile crowdsensing,

C. Tu, Z. Yu, J. Huang, F. Huang, Y . Wu, L. Han, L. Wang, and R. Huang, “Adaptive role learning with evolutionary multi-agent re- inforcement learning for uav-vehicle collaboration in sparse mobile crowdsensing,”IEEE Internet of Things Journal, 2025

2025

[4] [1]

A robust trust management system for v2x networks integrating isac with blockchain smart contracts,

M. U. F. Qaisar, W. Yuan, L. Zhang, S. A. Chaudhry, G. Han, and Y . Zhang, “A robust trust management system for v2x networks integrating isac with blockchain smart contracts,”IEEE Transactions on Network and Service Management, vol. 23, pp. 2428–2439, 2026

2026

[5] [2]

The role of isac in 6g networks: Enabling next-generation wireless systems,

M. U. F. Qaisar, W. Yuan, O. G ¨unl¨u, T. Riihonen, Y . Cui, L. Zhang, N. Gonz ´alez-Prelcic, M. di Renzo, and Z. Han, “The role of isac in 6g networks: Enabling next-generation wireless systems,”IEEE Transactions on Network Science and Engineering, vol. 13, pp. 7825– 7861, 2025

2025

[6] [3]

Low-altitude wireless networks: A comprehensive survey,

W. Jun, Y . Yaoqi, Y . Weijie, L. Wenchao, W. Jiacheng, M. Tianqi, Z. Lin, C. Yuanhao, L. Fan, S. Genget al., “Low-altitude wireless networks: A comprehensive survey,”China Communications, vol. 23, no. 3, pp. 99–141, 2026

2026

[7] [4]

Mobile edge computing via a uav- mounted cloudlet: Optimization of bit allocation and path planning,

S. Jeong, O. Simeone, and J. Kang, “Mobile edge computing via a uav- mounted cloudlet: Optimization of bit allocation and path planning,” IEEE Transactions on Vehicular Technology, vol. 67, no. 3, pp. 2049– 2063, 2017

[8] [5]

Caching on the sky: A multiagent federated reinforcement learning approach for uav-assisted edge caching,

X. Li, J. Liu, X. Chen, J. Wang, and M. Pan, “Caching on the sky: A multiagent federated reinforcement learning approach for uav-assisted edge caching,”IEEE Internet of Things Journal, vol. 11, no. 17, pp. 28 213–28 226, 2024

2024