{"id":"4ce28d8f-48f1-457b-81ef-ba15a5171d0d","arxiv_id":"2411.18129","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"A two-layer algorithm with Gibbs-sampling task placement, greedy subcarrier allocation, and closed-form computing allocation reduces CAV sensing completion time in CARLA while meeting an accuracy constraint.","lead":"This paper proposes an edge-assisted cooperative sensing scheme in which roadside units and connected autonomous vehicles share LiDAR data and computing resources, and decide for each sensing task whether to compute it locally or at the roadside unit. The goal is to reduce total task completion time while holding sensing accuracy above a threshold; CARLA simulations show the proposed algorithm lowers completion time versus benchmark schemes.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Accuracy surrogate used as both constraint and evaluation metric leaves the sensing accuracy guarantee unverified.","rationale":"The reader's weakest_assumption identifies exactly the circular use of the DNN accuracy surrogate in constraint (5) and in the evaluation plot Fig. 2(b). This is the most load-bearing concern because the central claim includes 'while satisfying sensing accuracy constraint'; if the surrogate is biased, the constraint is meaningless and the claimed guarantee cannot be verified from the presented simulations. The reader's verdict CONDITIONAL is appropriate: the optimization structure is plausible and the completion-time reduction may hold, but the accuracy feasibility claim needs independent validation. I agree with the reader's assessment and propose a concrete check that would settle whether the concern lands. No further changes to the verdict are needed; the condition should be that the accuracy model is validated against true SECOND accuracy on held-out scenarios.","tokens_in":9395,"tokens_out":2715,"duration_ms":26789,"concrete_test":"Generate new CARLA scenarios with topologies, object sizes, and sensor views not represented in the 7,350-sample training set. For a set of fused point clouds, compute the true accuracy from the SECOND model's output probability for the true class. Compare these true values against predictions from f. Then take the placement and allocation decisions returned by Algorithm 2 for those scenarios and recompute the per-CAV average accuracy using the true SECOND accuracy rather than f. If any CAV falls below the A=0.85 threshold, or if the mean absolute error of f on the new scenarios exceeds a predefined bound (e.g., 0.05), the sensing accuracy guarantee is not established.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that the proposed scheme minimizes task completion time while satisfying the sensing accuracy constraint (5). This constraint uses a DNN fit a_{k,m}=f(q', lx, ly, lz) from Section II-F to predict the classification accuracy of fused point clouds. The same fitted function is then used in Fig. 2(b) as the accuracy metric when reporting the 38.7% accuracy improvement for CAV 1. In Section VI the authors state that a precise mathematical accuracy model is left for future work, effectively acknowledging that f is an empirical surrogate. If f deviates from the true accuracy of the SECOND detector on the deployed scenarios, then (i) the solution produced by Algorithm 2 may violate the actual accuracy threshold even though (5) is satisfied against f, and (ii) the accuracy improvements in Fig. 2(b) are measured with the same potentially biased model, making the 'satisfying accuracy constraint' part of the central claim circular. The reported test MSE of 0.01 is an average over a held-out split of the same synthetic data collection; it does not establish generalization across scenario distributions, object geometries, or sensor configurations that the RSU-side controller would encounter. Because the accuracy constraint is the only mechanism in P0 that couples task placement to sensing quality, an invalidated surrogate directly undermines the feasibility guarantee of the proposed solution.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper proposes an RSU-assisted cooperative sensing scheme for connected autonomous vehicles. The RSU shares its LiDAR data with CAVs; for each object within a CAV's region of interest, the CAV decides whether to classify locally (fusing its own point cloud with the RSU's broadcast data) or to offload to the RSU, where data from several CAVs may be fused. The authors formulate a mixed-integer nonlinear program (P0) that minimizes the total task completion time over task placement, subcarrier allocation, and computing-resource allocation, subject to a per-CAV average sensing accuracy constraint. They then decouple P0: a Gibbs-sampling outer layer chooses task placements, a greedy inner layer allocates subcarriers, and a closed-form Karush-Kuhn-Tucker solution allocates computing resources. Simulations in CARLA, using the SECOND detector to train a DNN accuracy estimator, show convergence behavior, accuracy above the 0.85 threshold, and reduced completion time compared with standalone sensing, random task placement, and equal resource allocation.","tokens_in":9636,"tokens_out":10400,"duration_ms":90406,"significance":"The paper addresses a relevant problem in edge-assisted cooperative sensing, and its decomposition into task placement and resource allocation is sensible. The closed-form computing allocation of Eq. (10) is a clean result, and the greedy subcarrier allocation is a plausible low-complexity heuristic. If the sensing accuracy model were validated against the actual detector, the proposed framework would provide a useful design tool for RSU-coordinated cooperative sensing. However, the current evaluation does not verify the central accuracy guarantee against the true classifier; the accuracy values used in the objective constraint and in Fig. 2(b) appear to come from the same fitted DNN surrogate. This makes the paper's main claimed guarantee unsubstantiated until the surrogate is checked.","major_comments":[{"comment":"The sensing accuracy constraint (5) uses the DNN estimator a_{k,m} = f(q', lx, ly, lz) from Section II-F. Figure 2(b) reports the sensing accuracy of the proposed scheme and of standalone sensing, but the text does not state whether these accuracy values are computed by running the SECOND detector on the fused point clouds or by evaluating f. If they are computed with f, then the claimed 38.7% accuracy improvement for CAV 1 and the claim that the proposed scheme satisfies the sensing accuracy requirement are assertions about the surrogate, not about the actual detector. The reported MSE of 0.01 in Section V is an average over a held-out split of the training data collection and does not establish accuracy on the evaluation scenarios. Because Section VI states that a precise mathematical accuracy model is left for future work, the feasibility guarantee of constraint (5) is currently unverified. Please validate f against SECOND's actual outputs for the scenarios used in Fig. 2(b), or explicitly report that the accuracy results are surrogate predictions and temper the corresponding claims.","section":"§II-F and §V, Fig. 2(b)"},{"comment":"The inner while loop in lines 4–7 randomly changes a single placement value until Constraint (5) is satisfied. The paper does not establish that a feasible placement exists for a given accuracy threshold A, nor that this random-walk procedure terminates; if no placement satisfies (5), the loop may run indefinitely. Please add an existence condition for the accuracy constraint or a termination bound with an infeasibility flag, and discuss how the algorithm behaves when A is set above the achievable accuracy.","section":"§IV-B, Algorithm 2"}],"minor_comments":[{"comment":"In the Lagrange function before Eq. (10), the term λ(∑_{m∈M_k^{comp}} v_{k,m} - 1) sums over the wrong index; constraint (4) is per computing node m, so the multiplier should be node-specific λ_m and the sum should be over k. The final formula is correct under the intended interpretation, but the notation should be fixed.","section":"§IV-A, Eq. (10)"},{"comment":"Line 1 says 'Randomly generate a feasible decision E' but no procedure is given, and line 14 refers to a stopping criterion without defining it. Please specify both.","section":"§IV-B, Algorithm 2"},{"comment":"The sentence 'for τ = 10^{-4}, the algorithm converges to the global optimal solution' is unsupported; the figure only shows the objective value over iterations. Please remove the global-optimality claim or provide a comparison with optimal solutions on small instances.","section":"§V, Fig. 2(a)"},{"comment":"The initialization assigns one subcarrier to every offloaded task, which requires the number of offloaded tasks to be no larger than B/B_s. Please state this assumption and address the case where it is violated.","section":"§IV-A, Algorithm 1"},{"comment":"The simulation results appear to be based on a single random seed; please report averages and confidence intervals over multiple random scenarios to support the comparisons in Fig. 2.","section":"§V"}],"recommendation":"major_revision","confidential_remarks":"The main concern is the circularity of the accuracy evaluation: the same DNN surrogate is used both in the optimization constraint and as the reported accuracy metric. This is fixable by validating the surrogate against SECOND on the evaluation scenarios or by clearly labeling the reported accuracy as surrogate-based. The optimization framework itself is sound and within the journal's scope."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nIf you pick up this paper, know this: the optimization core is solid and worth reading, but the sensing-accuracy part is weaker than it looks. The authors formulate a joint task placement and resource allocation problem for RSU-assisted cooperative sensing, decompose it into placement and allocation subproblems, and solve them with a Gibbs-sampling outer loop, a greedy subcarrier allocation, and a closed-form KKT water-filling solution for computing resources. The KKT derivation is correct, and the decomposition is sensible. CARLA-based simulations show meaningful completion-time reductions, especially when CAV computing resources are limited.\n\nWhat is actually new is the specific combination: accuracy-aware sensing from Ye et al. [7] is embedded into a joint placement/resource-allocation optimization, and the two-layer Gibbs-sampling solver is new in this setting. The building blocks are standard, but the formulation and solver together are a reasonable contribution to the vehicular edge computing literature.\n\nThe soft spots are real and concentrated in the accuracy model. Constraint (5) uses a DNN fit f(q', lx, ly, lz) to predict classification accuracy from data quality. The same fitted function is then used to report the accuracy improvements in Fig. 2(b), including the 38.7% gain for CAV 1. The paper itself admits in the conclusion that a precise mathematical accuracy model is left for future work, which confirms that the accuracy constraint is an empirical surrogate rather than a validated law. The reported test MSE of 0.01 is over a held-out split of the same synthetic data; it does not establish generalization across scenario distributions, object geometries, or sensor configurations. If the surrogate is biased, the claimed feasible solutions may violate the true accuracy threshold, and the accuracy gains in Fig. 2(b) are circular rather than externally verified. This does not invalidate the optimization framework, but it undermines the claim that the scheme \"satisfies the sensing accuracy requirement\" in any verifiable sense.\n\nTwo smaller issues: the global optimality claim for tau = 1e-4 is empirical, not proven — Gibbs sampling with a temperature parameter does not generally guarantee global optimality. And there is no code or dataset released, so reproducibility is limited to the paper's description.\n\nWho is this for? Readers working on vehicular edge computing, cooperative perception, and task offloading will get value from the problem formulation and the decomposition approach. It deserves a serious referee, but the referee should push for external validation of the accuracy surrogate, error bars across random scenarios, and ideally code release.\n\nRecommendation: send it to review with a request for major revision focused on validating the accuracy model against the actual SECOND detector on a held-out set of scenarios, not just a split of the training distribution.","headline":"A credible optimization framework for RSU-assisted cooperative sensing, but the accuracy guarantee rests on an unvalidated DNN surrogate that is also used to measure accuracy gains.","tokens_in":10156,"tokens_out":2518,"would_cite":false,"duration_ms":22346,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"RSU-assisted cooperative LiDAR sensing cuts CAV task completion time while holding accuracy above a threshold.","keywords":["edge-assisted cooperative sensing","task placement","resource allocation","Gibbs sampling","connected autonomous vehicles","LiDAR point cloud fusion","completion time minimization","sensing accuracy constraint"],"falsifier":"Run the same CARLA setup but replace the fitted DNN accuracy function with the actual SECOND classification accuracy computed on each fused point cloud, then re-run the two-layer algorithm; if the resulting allocation falls below the 0.85 average-accuracy threshold or the reported improvement over standalone sensing disappears, the central claim of accuracy-constrained completion-time minimization fails.","tokens_in":9195,"feed_emoji":"🚗","tokens_out":4331,"duration_ms":37919,"temperature":0.7,"pith_summary":"This paper proposes that a roadside unit (RSU) with its own LiDAR and edge server can help connected autonomous vehicles fuse point-cloud data and choose where to compute each sensing task, reducing the total completion time while keeping classification accuracy above a threshold. The authors formulate the joint task placement, subcarrier allocation, and computing resource allocation as a mixed-integer nonlinear program and split it into an outer Gibbs-sampling layer for placement and an inner layer with a greedy subcarrier assignment plus a closed-form convex resource allocation. In CARLA simulations with the SECOND detector, the scheme meets the 0.85 accuracy target and lowers completion time relative to standalone sensing, random placement, and equal resource allocation. A sympathetic reading is that cooperative RSU-assisted sensing is a practical way to accelerate perception tasks when onboard computing is limited.","feed_headline":"RSU-assisted LiDAR fusion cuts CAV sensing time","feed_subtitle":"A two-layer algorithm picks where each sensing task runs and how to split spectrum and computing to beat standalone sensing.","key_machinery":"The load-bearing object is the closed-form computing-resource allocation $v^\\star_{k,m}$ derived from the convex subproblem P3 via Karush–Kuhn–Tucker conditions: each task on a node receives a computing fraction proportional to $\\sqrt{\\eta_{k,m}C_{k,m}}$, where $\\eta_{k,m}$ counts how many CAVs depend on that task's completion and $C_{k,m}$ is the CPU demand. Around it, the outer Gibbs sampler proposes single-task placement flips and accepts them with probability $1/(1+e^{(\\hat\\Delta-\\Delta)/\\tau})$, while the greedy subcarrier algorithm iteratively awards each next subcarrier to the transmission that most reduces the data-upload time. The accuracy constraint is enforced by a DNN surrogate $a_{k,m}=f(q'_{k,m}, l^x_k, l^y_k, l^z_k)$ fitted offline to simulated quality-accuracy pairs; this surrogate converts a non-analytic perception-performance relation into a constraint the optimizer can evaluate.","core_discovery":"The central claim is that selectively fusing RSU and CAV LiDAR data and jointly deciding task placement and resource allocation minimizes total sensing-task completion time subject to an average accuracy constraint. The paper models each object classification as a task that can run locally on the CAV or at the RSU; if it runs at the RSU, bandwidth must be allocated for upload and the RSU waits for all contributing CAVs. The solution decouples the problem: Gibbs sampling explores task-placement decisions, a greedy algorithm assigns subcarriers to transmission tasks, and a closed-form expression $v^\\star_{k,m} = \\sqrt{\\eta_{k,m}C_{k,m}}/\\sum_k \\sqrt{\\eta_{k,m}C_{k,m}}$ allocates computing fractions proportionally to the square root of demand. The authors report that the two-layer algorithm converges, satisfies the accuracy threshold, and achieves the lowest completion time among benchmark schemes, with the largest gains when CAV computing resources are scarce.","pith_inferences":["If the accuracy surrogate were replaced by a certified bound rather than a learned fit, the same optimization framework would yield hard accuracy guarantees instead of empirical ones.","The square-root allocation rule $v^\\star_{k,m} \\propto \\sqrt{\\eta_{k,m}C_{k,m}}$ suggests a general edge-offloading heuristic: prioritize tasks that many vehicles depend on and that have heavy computing demand.","Because transmission time is governed by the slowest contributing CAV, the scheme implicitly rewards grouping tasks by upload bottleneck; a natural extension is jointly selecting which CAV views to fuse, not only where to place the task.","The Gibbs placement layer could be warm-started across consecutive frames as objects enter and leave the RSU coverage, turning the static optimization into an online perception pipeline."],"forward_implications":["If the central claim holds, RSU-assisted cooperative sensing can cut task completion time, especially when CAV onboard computing is weak.","The two-layer structure makes the mixed-integer problem tractable: a Gibbs-sampling outer layer plus convex inner subroutines yields a practical algorithm for joint placement and allocation.","The closed-form resource allocation lets computing fractions be recomputed instantly for any placement, enabling fast evaluation of many placement candidates.","In the simulated CARLA setup, the scheme keeps every CAV above the 0.85 average-accuracy requirement, with CAV 1 improving 38.7% over standalone sensing.","The convergence plots indicate that the algorithm converges fastest at temperature $\\tau = 10^{-4}$ in the tested instance."],"supporting_citations":[{"why":"Defines the data-quality descriptor $q_{k,m}$ and the DNN accuracy-estimation approach $f$ that the paper adopts, and supplies the accuracy-aware cooperative sensing baseline.","marker":"[7]"},{"why":"Supplies SECOND, the sparsely embedded convolutional detector used as the AI model for object classification and accuracy evaluation.","marker":"[13]"},{"why":"Provides the global coordinate transformation used to align point clouds from different sensors before fusion.","marker":"[14]"},{"why":"Supports the premise that infrastructure-assisted cooperative perception improves 3D object detection in driving scenarios.","marker":"[4]"},{"why":"Motivates the heterogeneous-computing assumption by documenting the large computing-capability gap between vehicles and edge nodes.","marker":"[12]"}],"fun_headline_variants":["Edge-assisted RSU-CAV LiDAR fusion cuts sensing delays","Resource allocation and task placement speed up CAV sensing","Cooperative sensing with Gibbs-sampled task placement","RSU-CAV cooperation trims task completion time","Two-layer algorithm lowers CAV sensing latency"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The whole accuracy guarantee rests on the fitted DNN surrogate relating fused point-cloud quality to classification accuracy; if that surrogate does not match the real detector's accuracy, the claimed 0.85 guarantee and 38.7% accuracy improvement are not established.","fun_headline_variants_meta":{"raw":{"variants":["Edge-assisted RSU-CAV LiDAR fusion cuts sensing delays","Resource allocation and task placement speed up CAV sensing","Cooperative sensing with Gibbs-sampled task placement","RSU-CAV cooperation trims task completion time","Two-layer algorithm lowers CAV sensing latency"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000238,"raw_usage":{"total_tokens":1502,"prompt_tokens":930,"completion_tokens":572,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":546,"completion_tokens_details":{"reasoning_tokens":496}},"tokens_in":546,"tokens_out":572,"duration_ms":5363,"temperature":1.0,"reasoning_tokens":496,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T11:29:42.666835+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the same CARLA setup but replace the fitted DNN accuracy function with the actual SECOND classification accuracy computed on each fused point cloud, then re-run the two-layer algorithm; if the resulting allocation falls below the 0.85 average-accuracy threshold or the reported improvement over standalone sensing disappears, the central claim of accuracy-constrained completion-time minimization fails.","supporting_citations":[{"cited_title":"Accuracy-aware cooperative sensing and computing for connected autonomous vehicles,","cited_arxiv_id":null,"evidence_quote":"Defines the data-quality descriptor $q_{k,m}$ and the DNN accuracy-estimation approach $f$ that the paper adopts, and supplies the accuracy-aware cooperative sensing baseline."},{"cited_title":"Second: Sparsely embedded convolutional detection,","cited_arxiv_id":null,"evidence_quote":"Supplies SECOND, the sparsely embedded convolutional detector used as the AI model for object classification and accuracy evaluation."},{"cited_title":"Model-assisted learning for adaptive cooperative perception of connected autonomous vehicles,","cited_arxiv_id":null,"evidence_quote":"Provides the global coordinate transformation used to align point clouds from different sensors before fusion."},{"cited_title":"Cooperative per- ception for 3D object detection in driving scenarios using infrastructure sensors,","cited_arxiv_id":null,"evidence_quote":"Supports the premise that infrastructure-assisted cooperative perception improves 3D object detection in driving scenarios."},{"cited_title":"Joint task offloading and resource allocation for vehicular edge computing based on V2I and V2V modes,","cited_arxiv_id":null,"evidence_quote":"Motivates the heterogeneous-computing assumption by documenting the large computing-capability gap between vehicles and edge nodes."}],"review_version":1}