Pith. sign in

REVIEW 4 major objections 6 minor 40 references

This paper claims that separating drone sensing dispatch from wind-aware speed control, trained in two stages, raises system profit by 20–51% over baseline schedulers in simulations on real delivery orders from two Chinese cities.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · deepseek-v4-flash

2026-08-01 14:02 UTC pith:FRFFCARH

load-bearing objection Sensible two-timescale RL framing for delivery-drone sensing, but the energy model in Eq. (2) makes energy effectively free, so the velocity controller's claimed gains—and the headline 20–47% improvements—cannot be trusted until that model is fixed. the 4 major comments →

arxiv 2607.18874 v1 pith:FRFFCARH submitted 2026-07-21 cs.LG cs.CY

Reinforcement Learning for Delivery Drone-Based Participatory Sensing in Dynamic Environments

Xin Ouyang , Songxin Lei , Xusen Guo , Yutian Jiang , Sijie Ruan , Yuxuan Liang This is my paper
classification cs.LG cs.CY MSC 68T0568T0790C40
keywords participatory sensingdelivery dronesurban sensingreinforcement learningtwo-timescale decision-makingwind-aware velocity controlspatial crowdsourcingUAV scheduling
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

The paper tries to establish that sensing tasks can be piggybacked onto commercial delivery drones profitably even when wind changes flight speed and energy use, if scheduling is split into two timescales. It claims that a macro-level dispatcher choosing which drone takes which sensing task, plus a micro-level controller adjusting each drone's speed against wind, trained separately rather than jointly, outperform greedy and flat reinforcement-learning baselines. In simulations driven by three months of real delivery orders from Hangzhou and Shanghai with hourly wind traces, the proposed TSRL framework reports average system-profit improvements of 20.1% and 46.6% over the strongest baseline, with per-setting gains up to 25.2% and 51.2%. The reason a sympathetic reader cares: if true, urban sensing can ride almost free on existing logistics flights, but only when low-level flight execution is treated as a controllable decision rather than a fixed cost.

Core claim

On its own terms, the discovery is that the joint delivery-plus-sensing scheduling problem (SensUAV) is not best solved as one combined optimization. The paper argues the right abstraction is a two-timescale decomposition: an upper policy that, once every K micro steps, first scores UAVs then picks a sensing task conditioned on the chosen UAV, and a lower policy that, every micro step, outputs a continuous velocity command informed by wind speed and direction. Because the wind-aware power model P_i = K1||v_i||^2 + K2||v_i − w_t||^3 makes travel cost depend on wind, velocity becomes a genuine decision variable. The evidence is a simulator built from real delivery orders and weather traces, wh

What carries the argument

The load-bearing object is the two-timescale policy decomposition plus the wind-aware power model. Equation (2) models drone power as blade profile energy K1||v_i||^2 plus parasite energy K2||v_i − w_t||^3, with constants taken from earlier UAV energy studies; this is what makes wind matter and gives the velocity controller something to optimize. The dispatcher policy is factorized as pi_h(a|s) = pi_h_u(u_i|s) * pi_h_z(z_j|s, u_i), so action selection is sequential: pick a UAV via masked softmax, then pick a sensing task conditioned on that UAV, with cross-attention and permutation-invariant pooling encoding UAV-task compatibility. Training is staged: first PPO learns the low-level velocity

Load-bearing premise

The entire advantage of the learned speed controller rests on the assumed power-consumption formula with fixed coefficients K1 and K2 being an accurate proxy for real drone energy under wind; the paper offers no calibration or sensitivity check against physical flights.

What would settle it

Fly the learned velocity controller on a physical drone with measured wind, logging energy per kilometer against a fixed-speed baseline and against the simulator's prediction; if the simulator's predicted energy savings vanish or reverse under real wind, the claimed profit improvements are simulator artifacts. A cheaper check is to re-run the experiments with K1 and K2 varied over plausible ranges and see whether TSRL's lead over fixed-speed baselines persists.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • Delivery fleets can double as urban sensing networks without dedicated sensing flights: the paper's simulator shows sensing detours and wind-aware speed adjustments raise system profit while preserving delivery completion.
  • Decoupling dispatch from velocity control, with the controller trained first, stabilizes learning; the ablation shows removing the velocity controller causes the largest profit drop, especially in the long-distance city.
  • Factorized UAV-then-task selection is faster than joint-action scoring: TSRL makes decisions in about 1.3 ms per step, well below the evaluated reinforcement-learning baselines.
  • In long-haul delivery settings, coordination quality matters more than fleet size; TSRL keeps improving as the fleet grows from 20 to 30 UAVs while several baselines plateau or degrade.

Where Pith is reading between the lines

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

  • If the wind-aware power model is approximately right, the same two-stage recipe should transfer to other logistics settings where an environmental field (wind, current, traffic, temperature) couples high-level task assignment to low-level motion cost — for example, marine delivery or electric ground fleets with congestion-dependent energy use.
  • The ablation result hints that velocity control may be the larger source of the gains; a direct test would be to keep the TSRL dispatcher but swap in a speed controller trained under a different wind model and measure how much of the profit advantage remains.
  • A testable extension is to feed short-term wind forecasts into the macro state so the dispatcher can anticipate energy costs instead of reacting to current wind only.
  • Because the shaping reward is used only during training, a cleaner deployment study could vary the shaping coefficient and the number of training steps together to separate reward-design effects from policy-capacity effects.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 6 minor

Summary. The paper studies SensUAV: a fleet of delivery drones is used to perform sensing tasks opportunistically during delivery, under time-varying wind conditions. The authors formulate an optimization problem whose objective (Eq. (3)) combines sensing entropy, delivery profit, and energy cost, with binary sensing assignments and continuous velocity decisions. They propose TSRL, a two-timescale reinforcement learning framework. A lower-level velocity controller is trained first with PPO and then frozen; an upper-level sensing dispatcher acts every K micro steps and factorizes each action by first selecting a UAV and then selecting a sensing task conditioned on that UAV. Experiments on LaDe Hangzhou/Shanghai datasets with Open-Meteo wind data report average system-profit improvements of 20.1% and 46.6% over the strongest baseline, with ablations and runtime measurements. The paper also sketches an NP-hardness proof, includes a case study, and promises code release.

Significance. If the empirical claims survive scrutiny, the paper addresses a relevant and underexplored problem: jointly coordinating delivery and participatory sensing under wind-dependent energy dynamics. The two-stage, two-timescale decomposition is a sensible way to handle the different temporal resolutions of dispatch and velocity control, and the factorized UAV-then-task dispatcher is a reasonable scalability measure. The use of real-world LaDe delivery data, real wind traces, ablations, and runtime reporting are strengths. However, the load-bearing conclusion that wind-aware velocity control drives the profit gains rests on the energy model in Eq. (2), and as instantiated that model makes energy essentially free; this must be corrected before the central claim can be accepted.

major comments (4)
  1. [§2.1, Eq. (2); §4.2] The energy model is numerically degenerate. With K1=1e-5 J·min²/m², K2=1e-9 J·min³/m³ and vmax=30 km/h=500 m/min, P_i≈2.5 J/min (wind term ~0.1 J/min). Against the 1000 kJ battery, endurance is ~3.8×10^5 min, so C4 is never binding, and w_e E is <10^-3 cents per flight while orders/tasks are 200/500 cents. Eq. (3) therefore contains no meaningful energy pressure, and the claim in §4.6 that TSRL's advantage comes from 'lower-level energy efficiency' is unsupported. Please recalibrate K1,K2 to realistic drone power (~100 W), state the actual units, and add sensitivity analysis; the w/o VS ablation and the wind-aware velocity controller must be re-evaluated under a non-degenerate energy model.
  2. [§3.3, Eq. (11); §3.4] The micro-level reward R^l is never specified. Eq. (11) defines the micro MDP but gives no reward; Stage I says only that PPO is used. Since the frozen low-level policy determines macro transitions and the w/o VS ablation is presented as the strongest component, the reader cannot tell whether learned velocities arise from energy minimization, delivery progress, or an unstated auxiliary objective. Specify the micro reward and its coefficients, and demonstrate that it aligns with Eq. (3).
  3. [§4.3, Table 2] No statistical significance tests are reported; the abstract's 'significantly outperforms' is not supported. Several entries have large relative standard deviations (e.g., Hangzhou nums=30 D2SN 2.74±0.72; Shanghai nums=20 TSRL 3.82±0.43). Report the number of seeds or test days, paired tests or confidence intervals, and the exact aggregation used for Table 2.
  4. [§4.2, baselines; §5] DeliverSense [6], Xiang et al. [33], and other UAV participatory-sensing methods are discussed in related work but not compared in Table 2. The RL baselines are adapted order-dispatching policies (SMORE, DECO, D2SN, DyPS) or simple rules; no closest prior UAV-sensing baseline is evaluated. If direct adaptation is infeasible, justify that in detail and include at least one direct comparison or a clearly stated limitation.
minor comments (6)
  1. [Definition 5, Eq. (2)] The phrase 'power consumption per minute' and the units of P_i are inconsistent; use W or J/min explicitly throughout.
  2. [Lemma 1] The NP-hardness proof is a sketch; formalize the knapsack capacity, item weights, and values, and account for the nonlinear entropy in Eq. (1).
  3. [§3.4] The sentence about 'achieves potential optimality' is vague and unsupported; please qualify or remove it.
  4. [§4.2] Sensing tasks are said to have time windows 'randomly sampled from 10 to 15 hours'; clarify whether this is a duration or an absolute interval and how expiration is handled.
  5. [Figures 4 and 5] Please define what error bars/whiskers represent and describe Figure 4's axes and variant legend in the text; currently the figure is referenced but not described.
  6. [References [8,37]] Verify that Eq. (2)'s coefficient values are indeed taken from those sources; the reported values do not correspond to typical drone power models.

Circularity Check

0 steps flagged

No significant circularity: the central profit result is a held-out empirical comparison, and the paper's self-citations are background, data, or baseline references, not load-bearing definitions.

full rationale

The paper is an empirical RL system paper. The central claim—that TSRL (hierarchical sensing dispatcher plus wind-aware velocity controller) improves system profit on real LaDe delivery data—is supported by held-out test-set evaluations (Table 2), with hyperparameters selected on validation (Sections 4.1 and 4.2). The reported 20.1% and 46.6% improvements are therefore not fitted into the result by construction. Eq. (2)'s energy model is taken from external references [8,37] and is not defined in terms of the claimed profit improvement; the 'w/o VS' ablation is an internal comparison, not a renamed prediction. Self-citations ([10] AgentSense for the coverage-metric alpha value, [15] HRL for background, [27] SMORE as a baseline method, and [31] LaDe as the dataset) serve as context, data sources, or comparison methods; none supplies a load-bearing theorem, uniqueness argument, or ansatz that forces the TSRL design. The limitation in Section 6 ('Future work will incorporate finer-grained aerodynamic modeling, and real-world deployment constraints') honestly flags the energy model as approximate; this is a calibration/validity risk, as the skeptic notes about Eq. (2)'s constants, but a calibration concern is not circularity. No definitional equivalence, fitted parameter disguised as a prediction, or self-citation chain is present.

Axiom & Free-Parameter Ledger

5 free parameters · 5 axioms · 0 invented entities

The central claim rests on the simulator's energy model and objective weights, which are imported from prior work without calibration, plus a two-stage training assumption that is asserted rather than compared against joint training. These are the main postulates the reader is asked to accept.

free parameters (5)
  • K1, K2 (energy model coefficients) = 1e-5 J·min²/m², 1e-9 J·min³/m³
    Set by hand following [8,37]; no calibration to real drone data. Central to wind-aware energy accounting.
  • Profit weights w_z, w_o, w_e = 2000, 200, 1e-5 cents/unit
    Chosen from 'industry prices and expert knowledge'; define the objective that all methods optimize, so they determine the reported profit deltas.
  • Entropy balance α = 0.5
    Set following [10]; affects the sensing coverage metric in Eq (1).
  • Shaping coefficient λ_shape = 2e-4
    Selected by hyperparameter study on validation (Fig 5a); affects training but not final evaluation.
  • Temporal abstraction K = 5 micro steps
    Selected by hyperparameter study (Fig 5b); controls the two-timescale horizon.
axioms (5)
  • domain assumption The power model P_i = K1||v||^2 + K2||v - w||^3 captures real drone energy consumption
    Invoked in Eq (2); no validation against physical drone measurements.
  • domain assumption The custom simulator faithfully reproduces urban delivery, weather traces, and sensing dynamics
    Experiments are entirely in the authors' simulator; transfer to real systems is assumed.
  • ad hoc to paper Two-stage training with a frozen low-level policy does not significantly harm optimality
    Section 3.4 argues non-stationarity justifies freezing, but no comparison to joint hierarchical training is provided.
  • ad hoc to paper Adapted order-dispatching baselines are fair representatives of prior UAV sensing methods
    Section 4.2 claims absence of literature, yet cites direct sensing baselines [6,17,33] and does not compare against them.
  • standard math PPO with MLP/attention policies converges to strong policies in this MDP
    Standard RL background; not proven for this specific non-stationary two-timescale setting.

pith-pipeline@v1.3.0-alltime-deepseek · 17342 in / 13651 out tokens · 119829 ms · 2026-08-01T14:02:44.035055+00:00 · methodology

0 comments
read the original abstract

Using Unmanned Aerial Vehicle (UAV) for urban sensing has emerged as a powerful paradigm to monitor the status of the city, e.g., air quality and noise levels, through agile aerial crowdsourcing. Despite this potential, existing UAV-based sensing approaches overlook environmental disturbances like wind that drastically impact drone velocity and energy efficiency. Consequently, directly applying existing methods to this joint delivery and sensing paradigm in dynamic environments faces two severe challenges: (1) scalability bottlenecks as fleet sizes expand; and (2) multi-timescale decision heterogeneity between macro task dispatching and micro velocity control. To tackle these, we formalize the problem as SensUAV and propose a Two TimeScale Reinforcement Learning framework (TSRL). Specifically, TSRL separates decision-making into two cooperative layers. At the macro level, a task-embedding sensing dispatcher handles scalability by separately encoding distinct task features and sequentially evaluating UAV suitability before task selection. At the micro level, a wind-aware velocity controller learns fine-grained velocity scheduling to adapt to dynamic environmental variations. Extensive experiments on real-world datasets demonstrate that TSRL significantly outperforms baselines, achieving average system profit improvements of 20.1% in Hangzhou and 46.6% in Shanghai.

Figures

Figures reproduced from arXiv: 2607.18874 by Sijie Ruan, Songxin Lei, Xin Ouyang, Xusen Guo, Yutian Jiang, Yuxuan Liang.

Figure 1
Figure 1. Figure 1: Participatory Sensing for Delivery Drones [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Framework for SensUAV 3.2 Sensing Task Dispatcher 3.2.1 MDP Formulation. We formulate the upper-layer sensing task dispatcher as a macro-level MDP: ℳℎ = (𝒮 ℎ , 𝒜ℎ , 𝒫 ℎ , ℛℎ , 𝛾 ), (5) where 𝒮 ℎ is the macro state space, 𝒜ℎ is the macro action space, 𝒫 ℎ is the macro transition function, ℛℎ is the macro reward function, and 𝛾 ∈ (0, 1) is the discount factor. Different from a standard single-timescale MDP, … view at source ↗
Figure 3
Figure 3. Figure 3: Sensing Dispatcher Policy Architecture The role of the velocity controller is to execute fine-grained flight control under dynamic urban conditions. In particular, once a way￾point or task target is determined by the task manager, the lower layer is responsible for adjusting UAV speed continuously so as to balance task execution and energy consumption. 3.4 Training Process As many stage-based RL frameworks… view at source ↗
Figure 4
Figure 4. Figure 4: Ablation study on system profit. lower system profit. Overall, TSRL provides a favorable trade-off between decision quality and inference efficiency, making it prac￾tical for real-time UAV sensing systems. 4.4 Ablation Study To assess the contribution of each component in TSRL, we conduct an ablation study by removing one design at a time. As shown in [PITH_FULL_IMAGE:figures/full_fig_p008_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Hyperparameter study on system profit. not only captures UAV-task compatibility more accurately but also significantly reduces computational overhead. 4.5 Hyperparameter Study Impact of 𝜆shape [PITH_FULL_IMAGE:figures/full_fig_p009_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: System behavior analysis. consequently sacrifices delivery completion, leading to lower de￾livery profit. In contrast, TSRL and DECO achieve higher delivery profit through more reasonable sensing dispatch, indicating that ef￾fective sensing coordination should not come at the expense of de￾livery service. The remaining profit gap between TSRL and DECO mainly comes from concurrent sensing allocation and low… view at source ↗
Figure 7
Figure 7. Figure 7: Case study. UAV 20 as free from delivery duties and dispatches it on a multi￾stop sensing tour, completing five sensing tasks located at the spa￾tial periphery of the depot. In contrast, DECO and SFG leave the same UAV stationary for most of the window, implicitly relying on UAVs already serving delivery orders to collect sensing data oppor￾tunistically. This strategy may systematically fail to cover perip… view at source ↗
Figure 8
Figure 8. Figure 8: Web Platform of SensUAV system To better understand how TSRL behaves after deployment, we build a lightweight visualization platform using the LaDe test set. The platform places the replayed UAV trajectories on an Open￾StreetMap background and shows how the agents move as sens￾ing and delivery demands change over time. We use ten test days from LaDe, so that the replay covers different demand patterns and … view at source ↗

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

40 extracted references · 4 linked inside Pith

  1. [1]

    Abdulwahab Ali Almazroi and Nasir Ayub. 2023. Multi-task learning for elec- tricity price forecasting and resource management in cloud based industrial IoT systems. IEEE Access 11 (2023), 54280–54295

    2023

  2. [2]

    Christopher Amato. 2024. An introduction to centralized training for decentral- izedexecutionincooperativemulti-agentreinforcementlearning. arXiv preprint arXiv:2409.03052 (2024)

    Pith/arXiv arXiv 2024

  3. [3]

    Dimitris Bertsimas and Bartolomeo Stellato. 2022. Online mixed-integer opti- mization in milliseconds. INFORMS Journal on Computing 34, 4 (2022), 2229– 2248

    2022

  4. [4]

    Jinwen Chen, Hao Miao, Dazhuo Qiu, Jiannan Guo, Yawen Li, and Yan Zhao

  5. [5]

    Xuecheng Chen, Haoyang Wang, Yuhan Cheng, Haohao Fu, Yuxuan Liu, Fan Dang, Yunhao Liu, Jinqiang Cui, and Xinlei Chen. 2024. Ddl: Empowering de- livery drones with large-scale urban sensing capability.IEEE Journal of Selected Topics in Signal Processing (2024)

    2024

  6. [6]

    Deliversense:Efficientdeliverydroneschedulingfor crowdsensing with deep reinforcement learning

    Xuecheng Chen, Haoyang Wang, Zuxin Li, Wenbo Ding, Fan Dang, Chengye Wu,andXinleiChen.2022. Deliversense:Efficientdeliverydroneschedulingfor crowdsensing with deep reinforcement learning. InAdjunct proceedings of the 2022 ACM international joint conference on pervasive and ubiquitous computing and the 2022 ACM international symposium on wearable computers...

    2022

  7. [7]

    Schedulingdrone and mobile charger via hybrid-action deep reinforcement learning.IEEE Trans- actions on Mobile Computing (2025)

    JizheDou,HaotianZhang,YangLuo,andGuodongSun.2025. Schedulingdrone and mobile charger via hybrid-action deep reinforcement learning.IEEE Trans- actions on Mobile Computing (2025)

    2025

  8. [8]

    Hao Gao, Wonjun Lee, Wuchen Li, Zhu Han, Stanley Osher, and H Vincent Poor. 2020. Energy-efficient velocity control for massive numbers of rotary- wing UAVs: A mean field game approach. InGLOBECOM 2020-2020 IEEE Global Communications Conference. IEEE, 1–6

    2020

  9. [9]

    Bin Guo, Yan Liu, Leye Wang, Victor OK Li, Jacqueline CK Lam, and Zhiwen Yu. 2018. Task allocation in spatial crowdsourcing: Current state and future directions. IEEE Internet of Things Journal 5, 3 (2018), 1749–1764

    2018

  10. [10]

    Xusen Guo, Mingxing Peng, Xixuan Hao, Xingchen Zou, Qiongyan Wang, Sijie Ruan, and Yuxuan Liang. 2026. AgentSense: LLMs empower generalizable and explainable web-based participatory urban sensing. InProceedings of the ACM Web Conference 2026. 5439–5450

    2026

  11. [11]

    Hierarchicalre- inforcementlearning:Asurveyandopenresearchchallenges

    MatthiasHutsebaut-Buysse,KevinMets,andStevenLatré.2022. Hierarchicalre- inforcementlearning:Asurveyandopenresearchchallenges. Machine Learning and Knowledge Extraction 4, 1 (2022), 172–221

    2022

  12. [12]

    Shenggong Ji, Yu Zheng, and Tianrui Li. 2016. Urban sensing based on human mobility. InProceedings of the 2016 ACM International Joint Conference on Perva- sive and Ubiquitous Computing. 1040–1051

    2016

  13. [13]

    Wen Ji, Ke Han, and Tao Liu. 2023. A survey of urban drive-by sensing: An optimization perspective. Sustainable Cities and Society 99 (2023), 104874

    2023

  14. [14]

    Zeqiang Lai, Ying Fu, and Jun Zhang. 2024. Hyperspectral image super resolu- tion with real unaligned RGB guidance.IEEE Transactions on Neural Networks and Learning Systems (2024)

    2024

  15. [15]

    Songxin Lei, Chunming Ma, Haomin Wen, Yexin Li, Lizhenghe Chen, Qianyu Yang, Fugee Tsung, Lei Chen, Sijie Ruan, and Yuxuan Liang. 2026. Hierarchical Reinforcement Learning for Cooperative Air-Ground Delivery in Urban System. arXiv preprint arXiv:2602.12913 (2026)

    arXiv 2026

  16. [16]

    Timothy P Lillicrap, Jonathan J Hunt, Alexander Pritzel, Nicolas Heess, Tom Erez, Yuval Tassa, David Silver, and Daan Wierstra. 2015. Continuous control with deep reinforcement learning. arXiv 2015.arXiv preprint arXiv:1509.02971 (2015)

    Pith/arXiv arXiv 2015

  17. [17]

    Delay-Sensitive GoodsDeliveryandIn-SituSensingUsingaMulti-TaskDrone

    BinLiu,WeiNi,RenPingLiu,YJayGuo,andHongboZhu.2025. Delay-Sensitive GoodsDeliveryandIn-SituSensingUsingaMulti-TaskDrone. IEEE Transactions on Mobile Computing (2025)

    2025

  18. [18]

    Yao Lu, Shuai Wang, Yu Yang, Hai Wang, Baoshen Guo, Desheng Zhang, Shuai Wang, and Tian He. 2024. DECO: Cooperative order dispatching for on-demand delivery with real-time encounter detection. InProceedings of the 33rd ACM In- ternational Conference on Information and Knowledge Management . 4734–4742

    2024

  19. [19]

    Deepreinforcementlearning-basedwind disturbance rejection control strategy for uav.Drones 8, 11 (2024), 632

    Qun Ma, Yibo Wu, Muhammad Usman Shoukat, Yukai Yan, Jun Wang, Long Yang,FuwuYan,andLirongYan.2024. Deepreinforcementlearning-basedwind disturbance rejection control strategy for uav.Drones 8, 11 (2024), 632

    2024

  20. [20]

    Asad Mahmood, Thang X Vu, Wali Ullah Khan, Symeon Chatzinotas, and Björn Ottersten. 2025. UAV-Assisted 5G Networks: Mobility-Aware 3D Trajectory Op- timization and Resource Allocation for Dynamic Environments. In2025 IEEE 102nd Vehicular Technology Conference (VTC2025-Fall). IEEE, 1–7

    2025

  21. [21]

    Hao Qin, Zhaozhou Wu, Yuanwei Liu, Xingqi Zhang, and Xinyue Zhang. 2025. Physics-Based Trajectory Design for Cellular-Connected UAVin Rainy Environ- ments Based on Deep Reinforcement Learning.IEEE Transactions on Intelligent Transportation Systems (2025)

    2025

  22. [22]

    Socialdrone:Anin- tegratedsocialmedia anddronesensingsystemfor reliabledisasterresponse.In IEEE INFOCOM 2020-IEEE Conference on Computer Communications

    MdTahmidRashid,DanielYueZhang,andDongWang.2020. Socialdrone:Anin- tegratedsocialmedia anddronesensingsystemfor reliabledisasterresponse.In IEEE INFOCOM 2020-IEEE Conference on Computer Communications . IEEE, 218– 227

    2020

  23. [23]

    John Schulman, Filip Wolski, Prafulla Dhariwal, Alec Radford, and Oleg Klimov

  24. [24]

    Hado Van Hasselt, Arthur Guez, and David Silver. 2016. Deep reinforcement learning with double q-learning. InProceedings of the AAAI conference on artifi- cial intelligence, Vol. 30

    2016

  25. [25]

    Jingwei Wang, Qianyue Hao, Wenzhen Huang, Xiaochen Fan, Zhentao Tang, Bin Wang, Jianye Hao, and Yong Li. 2024. Dyps: Dynamic parameter sharing in multi-agent reinforcement learning for spatio-temporal resource allocation. In Proceedings of the 30th ACM SIGKDD conference on knowledge discovery and data mining. 3128–3139

    2024

  26. [26]

    Learningtocut via hierarchical sequence/set model for efficient mixed-integer programming

    Jie Wang, Zhihai Wang, Xijun Li, Yufei Kuang, Zhihao Shi, Fangzhou Zhu, MingxuanYuan,JiaZeng,YongdongZhang,andFengWu.2024. Learningtocut via hierarchical sequence/set model for efficient mixed-integer programming. IEEE Transactions on Pattern Analysis and Machine Intelligence (2024)

    2024

  27. [27]

    Shuliang Wang, Song Tang, Sijie Ruan, Cheng Long, Yuxuan Liang, Qi Li, Ziqiang Yuan, Jie Bao, and Yu Zheng. 2024. Urban sensing for multi-destination workers via deep reinforcement learning. In2024 IEEE 40th International Confer- ence on Data Engineering (ICDE). IEEE, 4167–4179

    2024

  28. [28]

    Yu Wang, Jingfei Wu, Xingyuan Hua, Chi Harold Liu, Guozheng Li, Jianxin Zhao, Ye Yuan, and Guoren Wang. 2023. Air-ground spatial crowdsourcing with UAV carriers by geometric graph convolutional multi-agent deep reinforcement learning. In2023 IEEE 39th International Conference on Data Engineering (ICDE) . IEEE, 1790–1802

    2023

  29. [29]

    Learningcutselectionformixed-integerlinear programming via hierarchical sequence model.arXiv preprint arXiv:2302.00244 (2023)

    Zhihai Wang, Xijun Li, Jie Wang, Yufei Kuang, Mingxuan Yuan, Jia Zeng, Yong- dongZhang,andFengWu.2023. Learningcutselectionformixed-integerlinear programming via hierarchical sequence model.arXiv preprint arXiv:2302.00244 (2023)

    Pith/arXiv arXiv 2023

  30. [30]

    Franz Wesselmann and Uwe Stuhl. 2012. Implementing cutting plane manage- ment and selection techniques. InTechnical Report. University of Paderborn

    2012

  31. [31]

    Lixia Wu, Haomin Wen, Haoyuan Hu, Xiaowei Mao, Yutong Xia, Ergang Shan, Jianbin Zheng, Junhong Lou, Yuxuan Liang, Liuqing Yang, et al. 2024. LaDe: The first comprehensive last-mile express dataset from industry. InProceedings of the 30th ACM SIGKDD Conference on Knowledge Discovery and Data Mining . 5991–6002

    2024

  32. [32]

    Chuankai Xiang, Yanfang Mo, Wei Liu, Zhibin Wu, and Lishuai Li. 2025. Path pool based transformer model in reinforcement framework for dynamic urban drone delivery problem.Transportation Research Part C: Emerging Technologies 177 (2025), 105165

    2025

  33. [33]

    Chaocan Xiang, Yanlin Zhou, Haipeng Dai, Yuben Qu, Suining He, Chao Chen, and Panlong Yang. 2021. Reusing delivery drones for urban crowdsensing.IEEE Transactions on Mobile Computing 22, 5 (2021), 2972–2988

    2021

  34. [34]

    Jing Xu, Fangwei Zhong, and Yizhou Wang. 2020. Learning multi-agent coordi- nation for enhancing target coverage in directional sensor networks.Advances in Neural Information Processing Systems 33 (2020), 10053–10064

    2020

  35. [35]

    Yanggang Xu, Jirong Zha, Jiyuan Ren, Xintao Jiang, Hongfei Zhang, and Xin- lei Chen. 2024. Scalable multi-agent reinforcement learning for effective uav scheduling in multi-hop emergency networks. InProceedings of the 30th Annual International Conference on Mobile Computing and Networking . 2028–2033

    2024

  36. [36]

    XinlangYue,YiranLiu,FangzhouShi,SihongLuo,ChenZhong,MinLu,andZhe Xu. 2024. An end-to-end reinforcement learning based approach for micro-view order-dispatching in ride-hailing. InProceedings of the 33rd ACM international conference on information and knowledge management . 5054–5061

    2024

  37. [37]

    Kuangyu Zheng, Zimo Ma, Mingyue Zhao, Zhuyang Zhou, Ziheng Zhang, and Yifeng Li. 2022. Joint efficient UAV trajectory and velocity optimization for IoT data collection using a new projection algorithm.Drones 6, 12 (2022), 376

    2022

  38. [38]

    2019.Urban computing

    Yu Zheng. 2019.Urban computing. MIT Press. Conference acronym ’XX, June 03–05, 2018, Woodstock, NY Ouyang et al. A Real-World Deployment Visualization Figure 8: Web Platform of SensUA V system To better understand how TSRL behaves after deployment, we build a lightweight visualization platform using the LaDe test set. The platform places the replayed UAV ...

    2019

  39. [2017]

    Proximal policy optimization algorithms.arXiv preprint arXiv:1707.06347 (2017)

    Pith/arXiv arXiv 2017

  40. [2025]

    In2025 IEEE 41st International Conference on Data Engineering (ICDE)

    Sustainability-Oriented Task Recommendation in Spatial Crowdsourc- ing. In2025 IEEE 41st International Conference on Data Engineering (ICDE). IEEE, 2712–2725