{"id":"1b4d30cc-0171-4422-99c3-0cfcb27c7450","arxiv_id":"2506.15868","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"high","formal_verification":"none","parameter_count":4,"one_line_summary":"CooperRisk couples V2X cooperative perception with scene-consistent multi-agent trajectory prediction to produce temporal risk maps that reduce planned conflicts by about half on the V2XPnP dataset.","lead":"The authors propose CooperRisk, a pipeline that combines cooperative V2X perception, multi-agent trajectory prediction, and risk map generation to quantify driving risk for planning. It reports a 44.35% reduction in planned conflict rate on a real-world V2X dataset, which matters if V2X systems are to deliver safety benefits beyond single-vehicle sensors.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Equation (19) is not the gradient of the MPC cost in Eq. (16), so the planner may not minimize the stated risk map and the reported 44.35% conflict-rate reduction is unsubstantiated.","rationale":"The reader's rejection is justified, and the reader's suspicion that the risk cost function is the weak point is directionally correct. However, the sharper and more load-bearing problem is in Eq. (19): the published gradient update is not the gradient of the stated MPC cost. The extra scalar factor J and the term Q delta (delta^T Q delta) instead of 2 Q delta mean the descent direction is systematically different from the true gradient, so the planner is not minimizing V_h as claimed. This flaw is independent of the undisclosed constants c0-c4; even if those constants were supplied and correctly calibrated, the planning result as described would still be invalid. The cooperative perception and scene-consistent prediction components are plausible, and the use of a real V2X dataset is a positive element, but the headline experimental claim is the 44.35% conflict-rate reduction produced by the risk-map MPC, and that claim rests on an invalid optimization step. A finite-difference check followed by a re-run with the correct gradient would settle the matter. Therefore the REJECT verdict stands, with no change.","tokens_in":11770,"tokens_out":7236,"duration_ms":73784,"concrete_test":"Implement the full cost J in Eq. (16) with the dynamics in Eqs. (13)-(15) and a fixed risk map V_h on a V2XPnP scenario. Compute the gradient of J with respect to U_k by finite differences and compare it with the analytical expression in Eq. (19) at several random (X,U) points. If they disagree, re-run the Table I planning comparison using the correct gradient, e.g., obtained by automatic differentiation of Eq. (16), and check whether the conflict-rate reduction from 0.08 to 0.04 is reproduced.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The most load-bearing defect is not only the undisclosed constants in Eq. (4); it is that the planning optimization as written is mathematically invalid. The cost in Eq. (16) is J = sum_k [ V_h(X_k) + (X_k - X_d^k)^T Q (X_k - X_d^k) ]. The update in Eqs. (18)-(19) defines grad_U J(U_k^i) = [dJ/dU]^T J(U_k^i) = (B_k^i)^T (dV_h/dX) V_h(X_k^i) + (B_k^i)^T Q delta (delta^T Q delta), where delta = X_k^i - X_d^{k,i}. This is not the gradient of Eq. (16): the first term would be the gradient of (1/2) V_h^2, not V_h, and the second term is Q delta (delta^T Q delta) rather than 2 Q delta, with the two terms scaled differently from the true gradient. Moreover, U_k affects X_{k+1} through Eq. (13), not X_k, so the same-index gradient ignores the causal structure of the dynamics. If the planner was actually implemented with Eq. (19), it is not minimizing the risk-map cost that the CR comparison claims to evaluate; if it was implemented with a correct gradient, then Eq. (19) misrepresents the method. Either way, the reported CR of 0.04 versus 0.08 cannot be attributed to the stated risk-map-based MPC, and this issue remains even if the constants c0-c4 were disclosed.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"CooperRisk is a V2X-enabled driving risk quantification pipeline that combines cooperative perception via late fusion, a transformer-based multi-agent multi-modal trajectory predictor, a risk map generator built from a hand-crafted potential function over predicted Gaussian trajectory distributions, and an MPC planner that uses the risk map as a cost objective. The paper claims to be the first V2X-enabled risk quantification pipeline covering perception through planning, and reports on the V2XPnP dataset a 44.35% reduction in conflict rate (CR 0.04 versus 0.08 for QCNet- and MTR-based pipelines), along with improvements in minADE, minFDE, trajectory overlap rate, and robustness to V2X noise.","tokens_in":12051,"tokens_out":6107,"duration_ms":60689,"significance":"The problem addressed is important: connecting cooperative perception, scene-consistent multi-agent prediction, and interpretable risk quantification for downstream planning is a valuable direction for safe autonomous driving. The scene-consistency argument is well motivated, and the evaluation on a real-world V2X dataset with ablations on fusion and noise robustness is a meaningful effort. If the planning results were backed by a correct and reproducible optimization, the pipeline would be a useful contribution. However, the current manuscript contains load-bearing mathematical and reproducibility issues that prevent the central claims from being accepted as stated.","major_comments":[{"comment":"Equation (19) is not the gradient of the cost in Eq. (16). The first term, (B_k^i)^T (∂V_h/∂X) V_h(X_k^i), is the gradient of (1/2)V_h^2 rather than of V_h, and the second term, (B_k^i)^T Q delta (delta^T Q delta), has an extra scalar factor (delta^T Q delta) and does not match the gradient 2Q delta of the quadratic term. Moreover, since X_{k+1}=A_k X_k + B_k U_k, the cost term V_h(X_k) depends on U_{k-1}, not on U_k, so the same-index chain rule in Eq. (19) is structurally inconsistent with the dynamics. Thus, if the planner was implemented with Eq. (19), it is not minimizing the stated risk-map cost; if it was implemented with a correct gradient, then Eq. (19) misrepresents the method. Either way, the reported conflict-rate reduction in Table I is unsubstantiated. The optimization is also underspecified: no constraints on U or X, no step size, no convergence criterion, and no feasibility discussion are provided.","section":"III.D, Eqs. (16)-(20)"},{"comment":"The five constants c0-c4 are described only as 'parameters with positive values' and their numerical settings and tuning procedure are never disclosed. Because the risk map V_h and hence the MPC cost in Eq. (16) directly depend on these constants, the CR numbers in Table I cannot be reproduced or audited. If the constants were selected using the evaluation scenarios, the reported 44.35% improvement could reflect tuning of the cost function rather than a benefit of V2X perception or scene-consistent prediction. The authors should disclose the values and include a sensitivity analysis.","section":"III.C, Eq. (4)"},{"comment":"Equation (10) is only a notation for a 'hierarchical distribution'; the actual functional form of V_h(s,l,t|...) is never defined. The text states that the risk map is obtained by 'mapping the risk distribution across the entire scene' and that 'the risk value is the quantity of the distribution,' which is not a computable definition. It is also unclear whether the planner uses the expected value over the predicted Gaussian modes, a quantile, or a worst-case value. Without an explicit formula relating V_h(X_k) to the predicted means, variances, covariances, and weights, Eq. (16) cannot be evaluated by a reader, and the experiments cannot be reproduced.","section":"III.C, Eq. (10) and 'Risk Map' paragraph"},{"comment":"The planning evaluation lacks statistical and procedural detail. No error bars, number of test scenarios, or significance tests are reported, and the computation of CR is not specified (e.g., how 'the probability of collision between the ego CAV's planned trajectory and the ground truth trajectories' is estimated). The headline 44.35% conflict-rate reduction rests on single point estimates, so its robustness and statistical significance are unknown. This is a major issue because the central claim of the paper is built on this comparison.","section":"IV.D, Table I and CR definition"}],"minor_comments":[{"comment":"The sentence 'We utilize the PointPillar network [33] as the LiDAR feature backbone as the is extracted' appears to contain a typo and should be rewritten.","section":"III.A"},{"comment":"Equation (4) is difficult to parse as printed; please clarify the placement of c2 and the exponent in the term involving ∆s.","section":"III.C, Eq. (4)"},{"comment":"The novelty claim 'to the best of our knowledge, this is the first driving risk quantification pipeline enhanced by V2X information' is strong; please cite and discuss the closest existing works on V2X-based risk assessment to support or qualify this claim.","section":"I and V"},{"comment":"The robustness results in Fig. 4 are reported without error bars or confidence intervals; please state whether these curves are single runs or averaged over multiple trials.","section":"IV.D, Fig. 4"},{"comment":"The EPA formula has a formatting issue: '|bNT P|' should presumably be '|N_TP|' or an equivalent notation.","section":"IV.A, Eq. (21)"}],"recommendation":"reject","confidential_remarks":"The central planning claim is invalid as presented because Eq. (19) does not compute the gradient of the stated cost, and the missing constants and vague risk-map definition prevent verification. This is not a local presentation issue: the reported CR improvement may not reflect the described method. I would not encourage a standard revision; however, if the authors can provide the correct optimization derivation, disclose all constants, and re-run the experiments with a clear specification of the planner, a fresh submission could be considered."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The integration itself is the new thing: late-fusion V2X perception feeding a multi-agent, multi-modal transformer predictor whose outputs become a risk map, which then drives an MPC planner. That is a real contribution to the connected-autonomy literature, and the prediction module's scene-consistency mechanism—pairing intention queries across agents—is a sensible idea that shows up in the TOR numbers. The paper also runs a noise/delay robustness study, which is more than most pipeline papers do.\n\nThe soft spots, however, are load-bearing. The MPC gradient in Eq. (19) is not the gradient of the cost in Eq. (16). For the dynamics in Eq. (13), U_k affects X_{k+1}, not X_k, so the first term should involve the chain through the next state, not V_h(X_k) directly. The second term has an extra (delta^T Q delta) factor. As printed, the optimizer is not minimizing the stated risk-map cost. If the implementation used a correct gradient, then the paper misrepresents the method; either way, the reported CR reduction (0.04 vs 0.08) cannot be attributed to the described planner. The five risk constants c0-c4 are never disclosed, only 'positive values,' so even the risk map itself is uncalibrated and unreproducible. There are also no error bars, no code/data release, and only one dataset. And the abstract's 44.35% reduction does not match Table I (0.08 to 0.04 is 50%). The prediction results are plausible—CooperRisk does beat the reimplemented QCNet/MTR baselines on minADE/minFDE—but without error bars or released code, I'd treat that as suggestive, not solid.\n\nOverall: this is a paper with a worthwhile idea and a flawed execution. The V2X-perception-to-risk-map framing is new and could influence later work, so it deserves a serious referee who can push for a corrected planner derivation, disclosed constants, and open artifacts. In its current form, the central planning claim is unsubstantiated, so I would not cite it and would not bring it to a reading group without a caveat.","headline":"V2X-perception-to-risk-map integration is genuinely new and the prediction module looks plausible, but Eq. (19) is not the gradient of Eq. (16), so the headline conflict-rate reduction is unsubstantiated as written.","tokens_in":12623,"tokens_out":5360,"would_cite":false,"duration_ms":49013,"reading_group":"maybe","serious_thinker":"no","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"A V2X-driven pipeline from cooperative perception to risk-map-based planning cuts the ego vehicle's conflict rate with background traffic by 44.35% on the V2XPnP dataset, the paper reports.","keywords":["V2X cooperative perception","driving risk quantification","risk map","multi-agent trajectory prediction","scene consistency","model predictive control","connected autonomous vehicles","trajectory overlap rate"],"falsifier":"Take the trained CooperRisk predictor and MPC planner, keep all weights fixed, and rerun the V2XPnP evaluation while perturbing each of $c_0$ through $c_4$ within a reasonable range; if the ego conflict rate rises above the 0.08 achieved by the QCNet- and MTR-based baselines under small perturbations, then the headline 0.04 result is carried by the hand-tuned potential, not by cooperative perception or scene-consistent prediction.","tokens_in":11535,"feed_emoji":"🚗","tokens_out":6544,"duration_ms":57689,"temperature":0.7,"pith_summary":"CooperRisk is presented as the first vehicle-to-everything (V2X) driving risk quantification pipeline that connects cooperative perception, multi-agent trajectory prediction, and risk-map-based motion planning. Its central claim is that fusing detections from multiple agents and predicting all agents' futures scene-consistently, instead of one object at a time, lets a planner quantify risk from both severity and exposure and lower the ego vehicle's conflict rate with background traffic by 44.35% on the real-world V2XPnP dataset (CR 0.04 versus 0.08 for QCNet- and MTR-based pipelines). This matters because single-vehicle perception suffers occlusion and limited range, and conflicting single-agent predictions tend to overestimate risk, narrow the drivable area, and make an autonomous vehicle hesitant or unsafe.","feed_headline":"V2X risk pipeline cuts driving conflict rate by 44.35 percent","feed_subtitle":"Cooperative perception and scene-consistent multi-agent prediction feed risk maps an MPC planner can avoid.","key_machinery":"The load-bearing object is the scenario risk map, a time-indexed field $V^h(s,l,t\\mid\\mu_{s,i,j,t},\\mu_{l,i,j,t},\\sigma_{s,i,j,t},\\sigma_{l,i,j,t},\\Sigma_{i,j,t},w_{i,j,t})$ produced by pushing each background agent's predicted Gaussian trajectory distribution through the two-object risk potential $V = c_0\\Delta v^2 + c_1 q / (c_2 \\Delta s\\, e^{c_3\\Delta v})^2 + c_4 \\Delta l^2$, where $\\Delta v$ is a mass-weighted relative speed (severity) and $\\Delta s,\\Delta l$ are heading-rotated longitudinal and lateral separations (exposure). Because the predictor outputs one Gaussian mixture per intention for every agent and enforces scene consistency through agent-level self-attention, cross-attention, and scene-level self-attention on paired intention queries, the resulting risk maps do not paint risks across trajectories that contradict each other. The MPC planner then minimizes the value of this risk map along its horizon together with a tracking cost, which is how the predicted distributions become a concrete conflict-rate reduction.","core_discovery":"The central claim is that a complete V2X pipeline can carry risk quantification all the way from raw shared sensor features to a planned trajectory, and that each stage improves the final conflict outcome. On V2XPnP, CooperRisk achieves the best prediction accuracy among the compared predictors (minADE 0.74 m and minFDE 1.26 m with V2X sharing), the lowest trajectory overlap rate (a 37.31% average reduction versus QCNet and MTR), and the lowest conflict rate in planning (CR 0.04). V2X sharing raises detection average precision from 51.0 to 58.0 AP@0.5 relative to single-agent perception, and using ground-truth perception lowers CR to 0.00, which the paper reads as evidence that upstream perception errors are the main residual source of conflict. The pipeline also keeps the best end-to-end perception-and-prediction accuracy under added V2X communication noise and delay.","pith_inferences":["The five constants $c_0$ through $c_4$ of the risk potential are never assigned numerical values; calibrating them on collision outcomes or learning them end-to-end would show whether the 44.35% conflict-rate reduction is robust or an artifact of the chosen potential.","Because the risk map is a time-indexed spatial field, it could be reused outside planning, for example as a real-time driver warning or as a safety monitor that flags imminent conflicts independent of the MPC.","A stricter ablation that fixes the prediction architecture and toggles only the paired-intention scene-consistency mechanism would isolate how much of the CR gain comes from scene consistency versus from the multi-modal GMM decoder itself.","The robustness curves suggest a testable prediction: if V2X noise increases beyond 1.0 m and 500 ms, the benefit of cooperative perception should eventually disappear; the paper does not report where that crossover lies."],"forward_implications":["V2X sharing alone raises detection AP by 12.07% (58.0 vs 51.0 at AP@0.5), so occluded and out-of-range objects that single-vehicle perception misses become visible to risk quantification.","Scene-consistent multi-agent multi-modal prediction cuts trajectory overlap rate by 37.31% on average, which prevents risk maps from flagging large parts of the drivable area as dangerous.","Risk-map-based MPC keeps the conflict rate below 9% even with single-modal predictors and reaches 0.04 with CooperRisk, so a distribution-based interpretable risk map can serve as the planner's cost without a learned reward.","With ground-truth perception the conflict rate drops to 0.00 for all multi-modal predictors, indicating that perception errors, rather than prediction errors, dominate the remaining conflicts."],"supporting_citations":[{"why":"Supplies the V2XPnP real-world sequential V2X dataset used for training, evaluation, and the communication noise/delay settings.","marker":"[22]"},{"why":"Provides the QCNet multi-modal trajectory prediction baseline that CooperRisk is compared against in Table I.","marker":"[35]"},{"why":"Provides the MTR multi-modal trajectory prediction baseline, the other 0.08 conflict-rate comparison point.","marker":"[28]"},{"why":"Motivates the scene-consistent prediction idea behind paired intentions and scene-level attention that avoid conflicting predicted trajectories.","marker":"[13]"},{"why":"Defines the end-to-end perception and prediction accuracy metric (EPA) used in the robustness assessment.","marker":"[36]"},{"why":"Supplies the late-fusion cooperative perception strategy and communication range assumptions adopted in the pipeline.","marker":"[32]"},{"why":"Provides the PointPillar LiDAR feature backbone used by each agent for detection.","marker":"[33]"},{"why":"Represents the prior integrated prediction and risk-quantification pipelines that CooperRisk extends by adding multi-agent multi-modal prediction.","marker":"[11]"}],"fun_headline_variants":["CooperRisk V2X fusion slashes conflicts by 44.35%","V2X risk maps from multi-agent prediction cut conflicts 44.35%","Interpretable V2X risk maps with cooperative prediction lower conflicts","Scene-consistent V2X prediction helps MPC planner avoid conflicts","CooperRisk pipeline: V2X perception + prediction reduces risk by 44.35%"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the hand-specified risk potential $V$ with its five unspecified 'positive' constants $c_0$ through $c_4$ is a faithful proxy for real collision risk; if those constants are miscalibrated or were tuned to the test scenarios, the reported 44.35% conflict-rate reduction could come from the cost function rather than from V2X perception or scene-consistent prediction.","fun_headline_variants_meta":{"raw":{"variants":["CooperRisk V2X fusion slashes conflicts by 44.35%","V2X risk maps from multi-agent prediction cut conflicts 44.35%","Interpretable V2X risk maps with cooperative prediction lower conflicts","Scene-consistent V2X prediction helps MPC planner avoid conflicts","CooperRisk pipeline: V2X perception + prediction reduces risk by 44.35%"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000608,"raw_usage":{"total_tokens":2853,"prompt_tokens":987,"completion_tokens":1866,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":603,"completion_tokens_details":{"reasoning_tokens":1767}},"tokens_in":603,"tokens_out":1866,"duration_ms":13685,"temperature":1.0,"reasoning_tokens":1767,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T19:30:58.471687+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take the trained CooperRisk predictor and MPC planner, keep all weights fixed, and rerun the V2XPnP evaluation while perturbing each of $c_0$ through $c_4$ within a reasonable range; if the ego conflict rate rises above the 0.08 achieved by the QCNet- and MTR-based baselines under small perturbations, then the headline 0.04 result is carried by the hand-tuned potential, not by cooperative perception or scene-consistent prediction.","supporting_citations":[{"cited_title":"Motion transformer with global intention localization and local movement refinement,","cited_arxiv_id":null,"evidence_quote":"Provides the MTR multi-modal trajectory prediction baseline, the other 0.08 conflict-rate comparison point."},{"cited_title":"Scept: Scene-consistent, policy- based trajectory predictions for planning,","cited_arxiv_id":null,"evidence_quote":"Motivates the scene-consistent prediction idea behind paired intentions and scene-level attention that avoid conflicting predicted trajectories."},{"cited_title":"Cooperfuse: A real-time cooperative perception fusion framework,","cited_arxiv_id":null,"evidence_quote":"Supplies the late-fusion cooperative perception strategy and communication range assumptions adopted in the pipeline."}],"review_version":1}