Pith. sign in

REVIEW 4 major objections 5 minor 1 cited by

Deep Reinforcement Learning-Based User Scheduling for Collaborative Perception

T0 review · 4 major / 5 minor · reviewed 2026-08-08 · deepseek-v4-flash

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

Pith's one-line read A deep reinforcement learning scheduler for V2X collaborative perception can be trained with a label-free reward based only on spatial confidence-map changes, and generally outperforms classic scheduling baselines.

desk verdict Useful label-free DRL scheduler for V2X collaborative perception, but the reward's equivalence to AP is asserted rather than proven, and the evaluation lacks error bars and comparisons to cited prior work. read the letter →

arxiv 2502.10456 v1 pith:DSC3IN2G submitted 2025-02-12 cs.LG cs.RO

Yandi Liu , Guowei Liu , Le Liang , Hao Ye , Chongtao Guo , Shi Jin This is my paper · ORCID
classification cs.LGcs.RO
keywords collaborativeperceptionV2XuserschedulingdeepreinforcementlearningdoubleQ-networkspatialconfidencemaplabel-freeobjectiveautonomousdriving
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 proposes SchedCP, a deep reinforcement learning scheduler that decides which nearby vehicles or roadside units may transmit their bird's-eye-view features to an ego vehicle during collaborative perception. The conventional goal of maximizing detection average precision needs costly ground-truth labels, so the paper replaces it with a label-free objective that rewards changes in the ego vehicle's spatial confidence map between consecutive scheduling slots. Training a double deep Q-network with that reward lets the scheduler use both instantaneous channel state and semantic confidence information to pick one link per slot. On the V2X-Sim dataset, the trained scheduler generally outperforms nearest-neighbor, round-robin, and maximum-rate baselines in AP@0.50 and AP@0.70, and a case study shows it resolves an occlusion scenario that the baselines miss.

What carries the argument

The spatial confidence map $\tau^t_e$ is the central object: each grid value estimates the probability that the corresponding region is occupied by an object. The label-free objective in Eq. (21a) counts grids where the classification confidence crosses the threshold $\zeta$ (term $T^t(x,y)$) and adds ReLU-activated squared confidence changes beyond a small threshold $\xi$ (term $G^t(x,y)$), so the reward in Eq. (22) is $R_t = \lambda_{r,nl} C^t_r + \lambda_u \sum_{(x,y)\in\Xi^t} \max\{T^t(x,y), G^t(x,y)\}$. A double deep Q-network with experience replay learns the Q-function from a state that includes channel gains, large-scale fading, and sums and maxima of squared remaining confidence maps. The feature selection rule $\mathrm{TOPK}(\tau^2_j (1-\tau^0_e) \mid B^t_j)$ decides which BEV grids each scheduled collaborator transmits, closing the loop from scheduling decisions to confidence-map updates.

What would settle it

Construct a V2X-Sim scene where a scheduled collaborator's feature shifts a detection box's position or orientation while the classification confidence map stays almost unchanged (for example, an object partially visible but mislocated), and compute the label-free utility in Eq. (21a): it stays near zero although AP@0.70 improves. If such cases occur frequently, reward (22) will not distinguish a perceptually useful link from a useless one.

Watch

Extended reading notes

Core claim

The central claim is that, for a well-behaved collaborative perception system, maximizing detection average precision can be replaced by maximizing the summed absolute shifts in the ego vehicle's spatial confidence map across scheduling slots. The paper formalizes this through Observations 3.1, 3.2, and 3.3: improving AP is proportional to reducing detection loss, which is dominated by classification-loss changes; predictions almost never flip from true to false; and a confidence change larger than a small threshold causes a beneficial prediction shift. This yields the label-free objective in Eq. (21a), summing $\max\{T^t(x,y), G^t(x,y)\}$ over transmitted grids, and the reward in Eq. (22) used to train a DDQN scheduler. The paper reports that SchedCP trained with this reward generally outperforms Nearest, round-robin, and Max Rate scheduling in AP@0.50 and AP@0.70 across 200–600 kHz bandwidths, and remains slightly below the labeled variant SchedCP-wl, especially at IoU 0.70 where localization losses matter.

Load-bearing premise

The load-bearing premise is Observation 3.1, that AP gains are always proportional to classification-confidence gains, so if a scheduled collaborator mainly tightens a bounding box or heading without changing confidence values the label-free reward is blind to the real perceptual benefit.

Editorial extensions

If this is right

  • Collaborative perception scheduling can be performed online without ground-truth labels: at decision time only CSI, confidence maps, and previous decisions are needed.
  • Because the reward includes transmission rate as a secondary term, the scheduler naturally trades perceptual gain against bandwidth usage.
  • The heuristic utility value deviates from AP in high-bandwidth regimes because collaboration there mainly reduces localization and direction loss rather than classification loss; the paper states this limitation explicitly.
  • In the occlusion case study on V2X-Sim, SchedCP detects all six target vehicles while nearest-neighbor, round-robin, and maximum-rate baselines detect only a subset.

Reading between the lines

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

  • The paper leaves implicit that its label-free objective is one-sided: gains that only tighten box localization or heading leave the classification confidence map unchanged, so a scheduled collaborator providing exactly that refinement would earn a near-zero reward even though AP@0.70 rises. A natural extension would add a label-free localization surrogate to reward (22).
  • The statistics supporting Observations 3.2 and 3.3 are measured on V2X-Sim scenes with few collaborators; in denser or adversarial occlusion scenes the threshold $\zeta$ likely needs re-estimation, and the under-1% violation rate should not be assumed to transfer.
  • Because the state transition includes the feature-selection rule $\Phi_{\mathrm{sel}}$, the same DDQN formulation can be retrained for other importance-map-based fusion backbones, suggesting the scheduler is a reusable policy layer on top of any confidence-producing perception stack.
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.

Referee Report

4 major / 5 minor

Summary. The paper studies V2X user scheduling for collaborative perception in autonomous driving. It formulates a sequential decision problem in which an ego vehicle selects one collaborator per scheduling slot to transmit BEV features, accounting for small-scale fading and limited bandwidth. Since perception labels are hard to obtain online, the authors propose a label-free reward based on changes in the ego vehicle's spatial confidence map, derived through three observations about the relationship between average precision, detection loss, and classification confidence. They train a DDQN agent (SchedCP) with either a label-dependent reward or the label-free reward and compare it against Nearest, Round Robin, and Max Rate baselines on the V2X-Sim dataset, reporting AP@0.50 and AP@0.70 across bandwidths. A case study illustrates how the learned scheduler balances channel state and semantic confidence.

Significance. If the label-free objective were validated, the paper would offer a practical method for real-time V2X scheduling without expensive annotation, using only CSI and spatial confidence maps. The problem setup is realistic, particularly the treatment of small-scale fading within a scheduling slot, and the case study gives useful intuition. The authors also make a genuine attempt to verify some of their assumptions empirically on V2X-Sim, and they compare two reward designs. However, the central theoretical chain from AP to the confidence-map utility is asserted rather than proved, and the paper's own results show the utility can disagree with AP in the high-bandwidth regime. The experimental evaluation also lacks error bars and multiple-seed statistics, so the claimed gains over baselines are not yet established with statistical confidence. The contribution is promising but needs substantial strengthening before it is convincing.

major comments (4)
  1. [Section III-C, Observation 3.1 and Eq. (21a)] The label-free objective rests on the unproved chain ΔAP ∝ ΔL_det ∝ max(ΔL_cls, ϱ). No proof is given beyond asserting that this 'occurs in most practical scenarios,' and the cited reference [33] only partially supports the AP-to-detection-loss link. Because (21a) maximizes changes in the confidence map τ, it is blind to improvements in box regression, orientation, or size that reduce L_loc and L_dir without changing τ. The paper should provide direct empirical evidence on V2X-Sim that ΔAP is monotonically related to ΔL_cls across scheduling decisions, or clearly restrict the claimed equivalence to scenarios where classification loss dominates.
  2. [Section III-C, Eq. (21a) and Eq. (22)] The threshold ξ used in G^t(x,y) = g(|τ^t_e(x,y)-τ^{t-1}_e(x,y)|^2 - ξ) and in the label-free reward (22) is never reported. Table I lists values of ζ but not ξ. Since the objective and reward depend critically on ξ, the value used in the experiments must be stated; otherwise the training procedure is not reproducible.
  3. [Section IV-B, Fig. 9 and accompanying text] The authors acknowledge that at bandwidths of 500-600 kHz the RR baseline has higher label-free utility than SchedCP-wl while its AP is no better, attributing this to the utility ignoring L_loc and L_dir. This is a direct admission that the label-free objective is not a faithful proxy for AP in the high-bandwidth regime. Since the paper's central claim is that (21a) approximates (10a), this inconsistency must be addressed, for example by modifying the objective to incorporate localization/direction information or by restricting the claim to low and moderate bandwidth regimes.
  4. [Section IV-B, Figures 7-10] No error bars or multiple-seed results are reported for the AP comparisons, so it is impossible to assess whether the claimed improvements over the baselines are statistically significant. At minimum, the authors should report the mean and standard deviation over several random seeds of the DDQN training and over the test frames.
minor comments (5)
  1. [Eq. (19)] The word 'otheriwse' should be 'otherwise'.
  2. [Eq. (7)] The index inside the logarithm appears as 'g t,ts_i' while the transmission rate is defined as C^{t,ts}_j; please check the subscript.
  3. [Section III-B] The symbol R_t is used both for the reward in (16) and (22) and for the selection mask R^t_j in (15); renaming one of them would avoid confusion.
  4. [Algorithm 1] The word 'paramater' should be 'parameter'.
  5. [Section III-C, Table I] The verification protocol for Observations 3.2 and 3.3 is only summarized; please explain how the True-to-False probability and the violation probability are computed over the dataset, and clarify whether the reported numbers are per-grid, per-frame, or aggregated.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity; the label-free reward is a heuristic proxy tested against an external AP benchmark, not an input refitted as a prediction.

full rationale

The derivation chain is not circular. The label-free objective in Eq. (21a) is proposed as a hand-designed heuristic proxy for AP via Observations 3.1 through 3.3, and the paper explicitly states that the consistency of (10a) and (21a) will be verified in Section IV rather than being assumed by construction. The DDQN agent maximizes reward (22) during training, but the central claim, that SchedCP generally outperforms the baselines, is evaluated against AP@0.50 and AP@0.70 on the held-out V2X-Sim test set, which is an external benchmark independent of the utility function. No equation in the paper reduces the claimed result to its inputs: the utility is not fitted to AP, and the reward weights and threshold zeta are selected as hyperparameters rather than as fitted predictors of the target metric. The self-citations (e.g., [17], [18], [29], [36]) are background citations for DRL-based resource allocation and channel modeling, and they are not load-bearing for the label-free objective or for the performance comparison. Observation 3.1 is asserted without full proof, and the paper itself admits in Section IV-B that the heuristic utility 'may deviate a little' from AP at high bandwidth because it ignores L_loc and L_dir; this is a correctness risk or limitation, not a circularity. The empirical calibration of zeta in Table I is a threshold check for Observation 3.3, not a fit of AP. Therefore the derivation is self-contained against external benchmarks and no circular step can be exhibited.

Assumptions & free parameters 5 free parameters · 6 assumptions · 0 invented entities

The central claim rests on several domain assumptions about the behavior of collaborative perception systems (Observations 3.1-3.3), on a single-link scheduling abstraction, and on hand-picked reward thresholds and weights. The label-free objective is a heuristic proxy, not a first-principles derivation. No new physical entities are introduced.

free parameters (5)
  • ζ (confidence threshold) = 0.05
    Set in Table I so that the True-to-False and violation probabilities are below 1%. Used in objective (20) and (21a).
  • ξ (change threshold)
    Introduced in Observation 3.3 as 'a small positive threshold' and used in the ReLU term G^t(x,y) in (21a) and the reward (22), but its value is never given. A load-bearing free parameter.
  • λ_r,l and λ_det = (0.02, 8)
    Reward weights for the label-dependent reward (16), chosen by hand in Table III.
  • λ_r,nl and λ_u = (0.04, 0.3)
    Reward weights for the label-free reward (22), chosen by hand in Table III.
  • Discount factor γ
    Used in the return formula (23) but not reported in Table III; an unspecified training hyperparameter.
assumptions (6)
  • domain assumption Observation 3.1: ∆APt ∝ ∆L_det ∝ max(∆L_cls, ϱ) and AP is nondecreasing for well-behaved collaborative perception systems
    Stated without proof in Section III-C. The entire label-free objective transformation depends on this proportionality. Citation [33] partially verifies the AP-to-loss link, but not the full chain in this setup.
  • domain assumption Observation 3.2: prediction never flips from true to false, only from false to true
    Empirically verified only on V2X-Sim in Table I, then generalized. Used to justify the indicator term T^t(x,y) in (20)-(21).
  • domain assumption Observation 3.3: when confidence moves away from the label, the change is bounded by √ξ
    Empirically verified only on V2X-Sim, and the ξ value is not specified. Used to replace the prediction-change indicator by a magnitude-maximization objective.
  • domain assumption The ego activates only one link per scheduling slot and all links share the same spectrum
    Stated in Section II-A as a simplification. This reduces the problem to selecting one of N links and avoids multi-link interference and scheduling.
  • domain assumption The spatial confidence map τ_j^0 can be transmitted to the ego with negligible overhead
    Stated in Section III-B: the overhead for transmitting τ_j^0 is negligible. This underpins the state design and the feasibility of the approach.
  • domain assumption The fusion network is a weighted sum with nonnegative weights proportional to confidence
    Section III-A, Eq. (12). The feature selection rule (13) relies on this form. The authors note DiscoNet, F-Cooper, and Where2comm meet this, but it is still a structural assumption.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Deep Reinforcement Learning-Based User Scheduling for Collaborative Perception." pith.science (2026). https://pith.science/paper/DSC3IN2G

@misc{pith2026250210456,
  author       = {Pith},
  title        = {Pith review of: Deep Reinforcement Learning-Based User Scheduling for Collaborative Perception},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/DSC3IN2G}},
  note         = {Machine review of arXiv:2502.10456}
}
read the original abstract

Stand-alone perception systems in autonomous driving suffer from limited sensing ranges and occlusions at extended distances, potentially resulting in catastrophic outcomes. To address this issue, collaborative perception is envisioned to improve perceptual accuracy by using vehicle-to-everything (V2X) communication to enable collaboration among connected and autonomous vehicles and roadside units. However, due to limited communication resources, it is impractical for all units to transmit sensing data such as point clouds or high-definition video. As a result, it is essential to optimize the scheduling of communication links to ensure efficient spectrum utilization for the exchange of perceptual data. In this work, we propose a deep reinforcement learning-based V2X user scheduling algorithm for collaborative perception. Given the challenges in acquiring perceptual labels, we reformulate the conventional label-dependent objective into a label-free goal, based on characteristics of 3D object detection. Incorporating both channel state information (CSI) and semantic information, we develop a double deep Q-Network (DDQN)-based user scheduling framework for collaborative perception, named SchedCP. Simulation results verify the effectiveness and robustness of SchedCP compared with traditional V2X scheduling methods. Finally, we present a case study to illustrate how our proposed algorithm adaptively modifies the scheduling decisions by taking both instantaneous CSI and perceptual semantics into account.

Figures

Figures reproduced from arXiv: 2502.10456 by the authors.

Figure 1
Figure 1. An illustrative example of collaborative perception. [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. An illustration of the collaborative perception framework for autonomous driving. [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Relationship among the sensor sampling interval, scheduling slot and [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figures from the paper (9 more)
Figure 4
Figure 4. Figure 4: Relationship between spatial confidence map and detection results. [PITH_FULL_IMAGE:figures/full_fig_p005_4.png]
Figure 5
Figure 5. Figure 5: Illustration of the scheduler as an agent interacting with the [PITH_FULL_IMAGE:figures/full_fig_p007_5.png]
Figure 6
Figure 6. Figure 6: Return on prepared validation set with increasing iterations. [PITH_FULL_IMAGE:figures/full_fig_p009_6.png]
Figure 7
Figure 7. Figure 7: Performance under different bandwidth conditions in terms of [PITH_FULL_IMAGE:figures/full_fig_p009_7.png]
Figure 9
Figure 9. Figure 9: The utility values of our methods compared with other baselines [PITH_FULL_IMAGE:figures/full_fig_p010_9.png]
Figure 8
Figure 8. Figure 8: Detection and classification loss under different bandwidth conditions. [PITH_FULL_IMAGE:figures/full_fig_p010_8.png]
Figure 11
Figure 11. Figure 11: Bird’s eye view of a scenario at the intersection. [PITH_FULL_IMAGE:figures/full_fig_p011_11.png]
Figure 10
Figure 10. Figure 10: Perception performance under different sensor sampling intervals in [PITH_FULL_IMAGE:figures/full_fig_p011_10.png]
Figure 13
Figure 13. Figure 13: Detection results of different scheduling strategies. [PITH_FULL_IMAGE:figures/full_fig_p012_13.png]

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. SComCP: Task-Oriented Semantic Communication for Collaborative Perception

    eess.SP 2025-07 conditional novelty 4.0 of 10

    SComCP combines importance-aware feature selection with a learned JSCC codec to improve collaborative 3D detection over noisy V2V channels, reporting gains at low SNR.

Reference graph

Works this paper leans on

40 extracted references · 32 canonical work pages · cited by 1 Pith paper

  1. [27]

    Accuracy-aware cooperative sensing and computing for connected autonomous vehicles,

    X. Ye, K. Qu, W. Zhuang, and X. Shen, “Accuracy-aware cooperative sensing and computing for connected autonomous vehicles,” IEEE Trans. Mobile Comput. , vol. 23, no. 8, pp. 8193–8207, Aug. 2024

    work page 2024

  2. [33]

    AP-loss for accurate one-stage object detection,

    K. Chen, W. Lin, J. Li, J. See, J. Wang, and J. Zou, “AP-loss for accurate one-stage object detection,” IEEE Trans. Pattern Anal. Mach. Intell. , vol. 43, no. 11, pp. 3782–3798, Nov. 2021

    work page 2021

  3. [1]

    Milestones in autonomous driving and intelligent vehicles: Survey of surveys,

    L. Chen, Y . Li, C. Huang, B. Li, Y . Xing, D. Tian, L. Li, Z. Hu, X. Na, Z. Li, S. Teng, C. Lv, J. Wang, D. Cao, N. Zheng, and F.-Y . Wang, “Milestones in autonomous driving and intelligent vehicles: Survey of surveys,” IEEE Trans. Intell. Veh. , vol. 8, no. 2, pp. 1046–1056, Feb. 2023

    work page 2023

  4. [2]

    Towards vehicle-to-everything au- tonomous driving: A survey on collaborative perception,

    S. Liu, C. Gao, Y . Chen, X. Peng, X. Kong, K. Wang, R. Xu, W. Jiang, H. Xiang, J. Ma et al. , “Towards vehicle-to-everything au- tonomous driving: A survey on collaborative perception,” arXiv preprint arXiv:2308.16714, 2023

    arXiv 2023

  5. [3]

    Collaborative perception in autonomous driving: Methods, datasets, and challenges,

    Y . Han, H. Zhang, H. Li, Y . Jin, C. Lang, and Y . Li, “Collaborative perception in autonomous driving: Methods, datasets, and challenges,” IEEE Intell. Transp. Syst. Mag. , vol. 15, no. 6, pp. 131–151, Nov.-Dec. 2023

    work page 2023

  6. [4]

    Vehicular communications: A physical layer perspective,

    L. Liang, H. Peng, G. Y . Li, and X. Shen, “Vehicular communications: A physical layer perspective,” IEEE Trans. Veh. Technol., vol. 66, no. 12, pp. 10 647–10 659, Dec. 2017

    work page 2017

  7. [5]

    Technical specification group radio access network; study on LTE-based V2X services; (Release 14),

    3rd Generation Partnership Project, “Technical specification group radio access network; study on LTE-based V2X services; (Release 14),” Jun. 2016

    work page 2016

  8. [6]

    Technical specification group radio access network; NR; study on NR vehicle-to-everything(V2X); (Release 16),

    3rd Generation Partnership Project, “Technical specification group radio access network; NR; study on NR vehicle-to-everything(V2X); (Release 16),” Mar. 2019

    work page 2019

Show all 40 references
  1. [7]

    Networking and communications in autonomous driving: A survey,

    J. Wang, J. Liu, and N. Kato, “Networking and communications in autonomous driving: A survey,” IEEE Commun. Surveys Tuts. , vol. 21, no. 2, pp. 1243–1274, Secondquarter 2019

    2019

  2. [8]

    Learning distilled collaboration graph for multi-agent perception,

    Y . Li, S. Ren, P. Wu, S. Chen, C. Feng, and W. Zhang, “Learning distilled collaboration graph for multi-agent perception,” in Proc. NIPS, 2021, pp. 29 541–29 552

    2021

  3. [9]

    V2X-ViT: Vehicle-to-everything cooperative perception with vision transformer,

    R. Xu, H. Xiang, Z. Tu, X. Xia, M.-H. Yang, and J. Ma, “V2X-ViT: Vehicle-to-everything cooperative perception with vision transformer,” in Proc. ECCV, 2022, pp. 107–124

    2022

  4. [10]

    V2VNet: Vehicle-to-vehicle communication for joint perception and prediction,

    T.-H. Wang, S. Manivasagam, M. Liang, B. Yang, W. Zeng, and R. Ur- tasun, “V2VNet: Vehicle-to-vehicle communication for joint perception and prediction,” in Proc. ECCV, 2020, pp. 605–621

    2020

  5. [11]

    Where2comm: Communication-efficient collaborative perception via spatial confidence maps,

    Y . Hu, S. Fang, Z. Lei, Y . Zhong, and S. Chen, “Where2comm: Communication-efficient collaborative perception via spatial confidence maps,” in Proc. NIPS, 2022, pp. 4874–4886

    2022

  6. [12]

    Technical specification group ser- vices and system aspects; enhancement of 3GPP support for V2X scenarios; stage 1 (Release 18),

    3rd Generation Partnership Project, “Technical specification group ser- vices and system aspects; enhancement of 3GPP support for V2X scenarios; stage 1 (Release 18),” Mar. 2024

    2024

  7. [13]

    Real-time spatio-temporal liDAR point cloud compression,

    Y . Feng, S. Liu, and Y . Zhu, “Real-time spatio-temporal liDAR point cloud compression,” in Proc. IROS, 2020, pp. 10 766–10 773

    2020

  8. [14]

    Low complexity outage optimal distributed channel allocation for vehicle-to-vehicle communi- cations,

    B. Bai, W. Chen, K. B. Letaief, and Z. Cao, “Low complexity outage optimal distributed channel allocation for vehicle-to-vehicle communi- cations,” IEEE J. Sel. Areas Commun. , vol. 29, no. 1, pp. 161–172, Jan. 2011

    2011

  9. [15]

    Cluster-based radio resource management for D2D-supported safety-critical V2X communi- cations,

    W. Sun, D. Yuan, E. G. Str ¨om, and F. Br ¨annstr¨om, “Cluster-based radio resource management for D2D-supported safety-critical V2X communi- cations,” IEEE Trans. Wireless Commun., vol. 15, no. 4, pp. 2756–2769, Apr. 2016

    2016

  10. [16]

    Resource allocation for D2D-enabled vehicular communications,

    L. Liang, G. Y . Li, and W. Xu, “Resource allocation for D2D-enabled vehicular communications,” IEEE Trans. Commun. , vol. 65, no. 7, pp. 3186–3197, Jul. 2017

    2017

  11. [17]

    Deep reinforcement learning based resource allocation for V2V communications,

    H. Ye, G. Y . Li, and B.-H. F. Juang, “Deep reinforcement learning based resource allocation for V2V communications,” IEEE Trans. Veh. Technol., vol. 68, no. 4, pp. 3163–3173, Apr. 2019

    2019

  12. [18]

    Spectrum sharing in vehicular networks based on multi-agent reinforcement learning,

    L. Liang, H. Ye, and G. Y . Li, “Spectrum sharing in vehicular networks based on multi-agent reinforcement learning,” IEEE J. Sel. Areas Com- mun., vol. 37, no. 10, pp. 2282–2292, Oct. 2019

    2019

  13. [19]

    Meta reinforcement learn- ing for fast spectrum sharing in vehicular networks,

    K. Huang, L. Liang, S. Jin, and G. Y . Li, “Meta reinforcement learn- ing for fast spectrum sharing in vehicular networks,” arXiv preprint arXiv:2309.17185, 2023

    2023 arXiv

  14. [20]

    Multi-agent deep reinforcement learning for dynamic power allocation in wireless networks,

    Y . S. Nasir and D. Guo, “Multi-agent deep reinforcement learning for dynamic power allocation in wireless networks,” IEEE J. Sel. Areas Commun., vol. 37, no. 10, pp. 2239–2250, Oct. 2019

    2019

  15. [21]

    AI empowered wireless communications: From bits to semantics,

    Z. Qin, L. Liang, Z. Wang, S. Jin, X. Tao, W. Tong, and G. Y . Li, “AI empowered wireless communications: From bits to semantics,” Proc. IEEE, vol. 112, no. 7, pp. 621–652, Jul. 2024

    2024

  16. [22]

    Semantic communica- tion for cooperative perception based on importance map,

    Y . Sheng, H. Ye, L. Liang, S. Jin, and G. Y . Li, “Semantic communica- tion for cooperative perception based on importance map,” J. Franklin Inst., vol. 361, no. 6, pp. 16–32, Mar. 2024

    2024

  17. [23]

    Autocast: Scalable infrastructure-less cooperative perception for dis- tributed collaborative driving,

    H. Qiu, P. Huang, N. Asavisanu, X. Liu, K. Psounis, and R. Govindan, “Autocast: Scalable infrastructure-less cooperative perception for dis- tributed collaborative driving,” arXiv preprint arXiv:2112.14947 , 2021

    2021 arXiv

  18. [24]

    MASS: Mobility-aware sensor scheduling of cooperative perception for connected automated driving,

    Y . Jia, R. Mao, Y . Sun, S. Zhou, and Z. Niu, “MASS: Mobility-aware sensor scheduling of cooperative perception for connected automated driving,” IEEE Trans. Veh. Technol., vol. 72, no. 11, pp. 14 962–14 977, Nov. 2023

    2023

  19. [25]

    C-MASS: Combinatorial mobility-aware sensor scheduling for collaborative per- ception with second-order topology approximation,

    Y . Jia, Y . Sun, R. Mao, Z. Nan, S. Zhou, and Z. Niu, “C-MASS: Combinatorial mobility-aware sensor scheduling for collaborative per- ception with second-order topology approximation,” arXiv preprint arXiv:2407.00412, 2024

    2024 arXiv

  20. [26]

    Select2Col: Leveraging spatial-temporal importance of se- mantic information for efficient collaborative perception,

    Y . Liu, Q. Huang, R. Li, X. Chen, Z. Zhao, S. Zhao, Y . Zhu, and H. Zhang, “Select2Col: Leveraging spatial-temporal importance of se- mantic information for efficient collaborative perception,” IEEE Trans. Veh. Technol., vol. 73, no. 9, pp. 12 556–12 569, Sept. 2024

    2024

  21. [28]

    Pointpillars: Fast encoders for object detection from point clouds,

    A. H. Lang, S. V ora, H. Caesar, L. Zhou, J. Yang, and O. Beijbom, “Pointpillars: Fast encoders for object detection from point clouds,” in Proc. CVPR, 2019, pp. 12 697–12 705

    2019

  22. [29]

    Spectrum and power allocation for vehicular communications with delayed CSI feedback,

    L. Liang, J. Kim, S. C. Jha, K. Sivanesan, and G. Y . Li, “Spectrum and power allocation for vehicular communications with delayed CSI feedback,” IEEE Wireless Commun. Lett. , vol. 6, no. 4, pp. 458–461, Aug. 2017

    2017

  23. [30]

    A comprehensive survey of LIDAR-based 3D object detection methods with deep learning for autonomous driving,

    G. Zamanakos, L. Tsochatzidis, A. Amanatiadis, and I. Pratikakis, “A comprehensive survey of LIDAR-based 3D object detection methods with deep learning for autonomous driving,” Comput. & Graph., vol. 99, pp. 153–181, Oct. 2021

    2021

  24. [31]

    F-cooper: Feature based cooperative perception for autonomous vehicle edge computing system using 3d point clouds,

    Q. Chen, X. Ma, S. Tang, J. Guo, Q. Yang, and S. Fu, “F-cooper: Feature based cooperative perception for autonomous vehicle edge computing system using 3d point clouds,” in Proc. of the 4th ACM/IEEE Symp. on Edge Comput., 2019, pp. 88–100

    2019

  25. [32]

    Prox- imal policy optimization algorithms,

    J. Schulman, F. Wolski, P. Dhariwal, A. Radford, and O. Klimov, “Prox- imal policy optimization algorithms,” arXiv preprint arXiv:1707.06347, 2017

    2017 arXiv

  26. [34]

    V2X-Sim: Multi-agent collaborative perception dataset and benchmark for autonomous driving,

    Y . Li, D. Ma, Z. An, Z. Wang, Y . Zhong, S. Chen, and C. Feng, “V2X-Sim: Multi-agent collaborative perception dataset and benchmark for autonomous driving,” IEEE Robot. Autom. Lett. , vol. 7, no. 4, pp. 10 914–10 921, Oct. 2022

    2022

  27. [35]

    Deep sparse rectifier neural networks,

    X. Glorot, A. Bordes, and Y . Bengio, “Deep sparse rectifier neural networks,” in Proc. JMLR Workshop Conf. , 2011, pp. 315–323

    2011

  28. [36]

    Deep-learning-based wireless resource allocation with application to vehicular networks,

    L. Liang, H. Ye, G. Yu, and G. Y . Li, “Deep-learning-based wireless resource allocation with application to vehicular networks,” Proc. IEEE, vol. 108, no. 2, pp. 341–356, Feb. 2020

    2020

  29. [37]

    Deep reinforcement learning with double Q-learning,

    H. Van Hasselt, A. Guez, and D. Silver, “Deep reinforcement learning with double Q-learning,” in Proc. AAAI, 2016, pp. 1–7

    2016

  30. [38]

    Human-level control through deep reinforcement learning,

    V . Mnih, K. Kavukcuoglu, D. Silver, A. A. Rusu, J. Veness, M. G. Bellemare, A. Graves, M. Riedmiller, A. K. Fidjeland, G. Ostrovski et al. , “Human-level control through deep reinforcement learning,” Nature, vol. 518, no. 7540, pp. 529–533, Feb. 2015

    2015

  31. [39]

    CARLA: An open urban driving simulator,

    A. Dosovitskiy, G. Ros, F. Codevilla, A. Lopez, and V . Koltun, “CARLA: An open urban driving simulator,” in Proc. Conf. Robot Learn., 2017, pp. 1–16

    2017

  32. [40]

    Microscopic traffic simulation using SUMO,

    P. A. Lopez, M. Behrisch, L. Bieker-Walz, J. Erdmann, Y .-P. Fl ¨otter¨od, R. Hilbrich, L. L ¨ucken, J. Rummel, P. Wagner, and E. Wießner, “Microscopic traffic simulation using SUMO,” in Proc. Conf. Robot Learn., 2018, pp. 2575–2582

    2018

Pith tools

Reviewed August 8, 2026 · model on record in the stance chip above.