{"id":"b43aa597-287d-42e5-b1a3-2b573254a305","arxiv_id":"2507.09505","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"TruckV2X is introduced as the first truck-centered cooperative perception dataset, providing multi-agent LiDAR and camera data from trailers, CAVs, and RSUs in CARLA simulation.","lead":"This paper presents TruckV2X, a simulated dataset of traffic scenarios captured from trucks, trailers, connected cars, and roadside sensors for cooperative perception research. It also runs object detection benchmarks and analyzes how trucks block other sensors' views.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Dataset scale claim is internally inconsistent: the 88,396-frame headline count does not match the 22,099-frame train/val/test split sum, leaving the 'large-scale' claim unverified.","rationale":"The reader's weakest assumption (sim-to-real fidelity) is legitimate and is explicitly conceded in Section VI, but it concerns the dataset's downstream value rather than the central existence/scale claim. The scale inconsistency is internal, checkable from the release, and directly attacks the 'large-scale' component of the claim. I would keep the conditional verdict: the paper is acceptable if the authors correct the frame-count definitions and reconcile the headline statistics with the split arithmetic; a documentation fix is sufficient, so no rejection is warranted. This does not change the reader's CONDITIONAL verdict, hence the adjustment is UNCHANGED.","tokens_in":12595,"tokens_out":11777,"duration_ms":126423,"concrete_test":"Download the released TruckV2X annotations and raw data from the HuggingFace link; count the number of unique synchronized timestamps per scenario in the train/val/test splits from the OpenCOOD annotation files, and count LiDAR point-cloud files and camera images. If total synchronized timestamps are 22,099, the abstract and Section III.D must be rewritten to state whether 'frames' mean per-agent frames or synchronized timestamps, and all scale-dependent claims and benchmark statistics must be recomputed and reported on the stated count. If the count is actually 88,396 synchronized timestamps, then the split table must be corrected and benchmark results rerun on the correct split.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is explicitly quantitative: TruckV2X is 'large-scale' with '88,396 frames of LiDAR point clouds, one million camera images, and 1.18 million 3D bounding box annotations' (Abstract). The paper never defines what counts as a frame, and the numbers do not reconcile. Section III.D splits the 64 scenarios into 38/9/17 scenarios with 13,243/2,839/6,017 frames; those sums are 22,099, exactly one quarter of 88,396. If 88,396 is a per-agent count (four agents per synchronized timestamp), then the synchronized corpus has 22,099 timestamps, and all benchmark tables (Tables III and IV) and the data statistics are based on the smaller synchronized set. The camera-image count is also unreconciled: with 10-15 cameras per timestamp, 22,099 timestamps yield ~220k-331k images, not one million, while 88,396 timestamps would require a different split arithmetic. Either a definitional error or a factor-of-four scale discrepancy is present in the paper. Since 'large-scale' is part of the novelty claim, this is a load-bearing correctness issue independent of the acknowledged sim-to-real limitation in Section VI.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces TruckV2X, a simulated (CARLA-based) dataset for truck-centered cooperative perception, claiming to be the first large-scale dataset of its kind. It includes synchronized multi-modal data from tractor, trailer, connected automated vehicle (CAV), and roadside unit (RSU) agents, an occlusion analysis that positions trucks as both occlusion sources and perception aids, and object-detection benchmarks using cooperative perception baselines. The authors provide a public Hugging Face repository and acknowledge the sim-to-real gap in the conclusion.","tokens_in":12860,"tokens_out":5344,"duration_ms":55248,"significance":"If the scale and definitions are corrected, TruckV2X would fill a genuine gap: existing V2X perception datasets focus on light vehicles, and existing truck datasets are single-agent. The paper's strengths include the publicly available dataset, reproduction of eight cooperative perception baselines within the OpenCOOD framework, and an explicit occlusion-recovery-rate analysis. These are concrete and falsifiable contributions. However, the central 'large-scale' claim is currently unverifiable because of an internal scale inconsistency, and the occlusion metrics are not precisely defined, which limits the current value until the authors address these points.","major_comments":[{"comment":"The reported dataset scale is internally inconsistent. The abstract states 88,396 frames of LiDAR point clouds and one million camera images, but the train/val/test split in Section III.D sums to 22,099 frames (13,243 + 2,839 + 6,017), exactly one quarter of the headline number. The camera count cannot be reconciled either: with 10–15 cameras per agent per timestamp, 22,099 synchronized timestamps yield roughly 220,000–330,000 images, not one million. The authors must define what constitutes a 'frame' (e.g., per-agent LiDAR frame versus synchronized multi-agent timestamp) and report both per-agent and synchronized totals consistently across the abstract, Section III.D, and Figure 5a.","section":"Abstract and Section III.D"},{"comment":"The occlusion recovery rate is a central quantitative contribution, but the paper never defines how an 'occluded object' is determined. Without a formal visibility model or an algorithm (e.g., raycasting against vehicle bounding boxes, line-of-sight checks, or a LiDAR point count criterion), the occlusion counts in Figures 6 and 8 and the ORR values are not reproducible. This is load-bearing because the paper's claim that trucks serve as 'mobile perception platforms' rests on these statistics.","section":"Section IV.B"},{"comment":"The benchmark setup conflates the truck ego with an early-fused tractor–trailer pair: Table III treats 'Truck' as a unified agent formed by early fusion, while Section III.A and Table IV treat tractor and trailer as separate cooperative agents. This makes the comparison between the two tables difficult to interpret, because the 'ego-only' truck baseline in Table III already includes a sensor combination that is itself a form of sensor fusion. Please state explicitly which sensor set constitutes the ego truck for each table and justify whether the early-fusion approach is an appropriate baseline.","section":"Section V.A and Table III"}],"minor_comments":[{"comment":"The labels 'CA Vs' and 'V2V and V2I' have inconsistent spacing and should read 'CAVs' and 'V2V/V2I' for clarity.","section":"Section II and Figure 2"},{"comment":"The LiDAR count is ambiguous: the text describes two LiDARs on the tractor and two additional LiDARs on the trailer (four total), while Table II lists '2× LiDAR' for the combined Tractor/Trailer row. Please clarify the per-agent LiDAR counts.","section":"Section III.B and Table II"},{"comment":"The phrase 'trained by 120 epochs' should be 'trained for 120 epochs'.","section":"Section V.A"},{"comment":"The title of reference [14] contains a typo: 'largs-scale' should be 'large-scale'.","section":"Reference [14]"}],"recommendation":"major_revision","confidential_remarks":"The paper has a solid contribution if the numbers are fixed. The scale inconsistency is likely a definitional issue rather than an intentional misstatement, but it currently makes the headline 'large-scale' claim unverifiable. I would request the authors to provide a clear frame definition, reconcile all counts, and add an explicit definition of occlusion for the ORR analysis. The absence of real-world validation is acknowledged in the paper and is acceptable for a simulation dataset, but the authors should temper the deployment-oriented claims in the abstract and conclusion."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nThe useful thing about this paper is the dataset itself, not the benchmarks. TruckV2X is a genuine first: a CARLA-based cooperative perception dataset with truck ego agents, separate trailer agents, a CAV, and an RSU, all with LiDAR and cameras. The trailer-as-agent idea is substantive—existing V2X datasets like OPV2V and DAIR-V2X are car-centric, and MAN TruckScenes, the main truck dataset, is single-agent and tractor-only. That gap is real, and the authors have filled it with a public artifact (HuggingFace link) organized in the OpenCOOD format. That alone makes the paper worth a serious look.\n\nThe occlusion analysis is a defined statistic, not a fitted parameter, and the ORR curves are an honest way to quantify how much cooperative agents recover occluded objects. The benchmarks cover five intermediate fusion methods plus early, late, and ego-only baselines, which is a competent if standard evaluation. The citation pattern is fine; the only self-citation (Traf-Align) is a related-work pointer, not a load-bearing reference.\n\nNow the soft spots, in proportion. The scale claim is the big one. The abstract says 88,396 frames and one million camera images. The train/val/test split sums to 22,099 synchronized frames (13,243+2,839+6,017). That is exactly a factor of four, so 88,396 is presumably the per-agent LiDAR frame count across the four agents. But with 15 cameras per synchronized timestamp, 22,099 frames gives roughly 330k images, not one million. Either the frame definition is inconsistent or the image count is wrong. Since 'large-scale' is part of the contribution, this needs to be fixed and defined before the paper is accepted.\n\nSecond, the benchmark tables have no error bars or multiple runs. Given that some method differences are around 2-3 AP points, you cannot tell if they are noise. The authors should report means and standard deviations over at least three seeds.\n\nThird, the sim-to-real gap. The authors admit in Section VI that the dataset lacks weather variation, sensor noise, and real-world validation. That is the right limitation to state, but the paper's value claims about truck-specific occlusions would be stronger with a fidelity study or at least a careful discussion of what CARLA does and does not model in trailer articulation.\n\nThe stress-test note about the scale discrepancy holds up on reading; it is not a manufactured issue. The rest of the paper is honest and the limitations are stated. I would send this to peer review. It needs a proper revision, but the dataset artifact and the trailer-agent concept deserve referee time.","headline":"A genuinely useful truck-centered cooperative perception dataset whose headline scale numbers do not add up; fix the frame definition and add error bars and it is a solid benchmark contribution.","tokens_in":13388,"tokens_out":5301,"would_cite":true,"duration_ms":48892,"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":"The paper introduces TruckV2X, the first large-scale truck-centered cooperative perception dataset, and uses it to show how tractor-trailer blind zones can be recovered through multi-agent collaboration.","keywords":["cooperative perception","autonomous trucking","articulated vehicle","LiDAR-camera dataset","occlusion analysis","vehicle-to-everything","simulation benchmark","tractor-trailer"],"falsifier":"A real-world field test that records a tractor-trailer, a connected car, and a roadside unit at the same intersection configurations and compares the measured occlusion recovery rates and occluded-object counts with TruckV2X's; a systematic mismatch would show the simulated benchmark does not represent real truck occlusions.","tokens_in":12412,"feed_emoji":"🚚","tokens_out":7403,"duration_ms":79837,"temperature":0.7,"pith_summary":"This paper claims that the absence of truck-specific cooperative perception data is a bottleneck for autonomous trucking, and introduces TruckV2X as the first large-scale dataset built around articulated trucks. The dataset is simulated, with synchronized LiDAR and camera streams from four agent types (tractor, trailer, connected car, and roadside unit) captured across 64 urban scenarios. Using it, the authors quantify how trailers create blind zones during turns, how trucks occlude other road users, and how much cooperative perception recovers. They also deliver detection benchmarks showing that cooperation consistently outperforms ego-only sensing. The claim matters because heavy-vehicle blind spots are a known safety problem, and cooperative perception needs a public testbed to develop solutions.","feed_headline":"TruckV2X is the first truck-centered cooperative perception dataset","feed_subtitle":"Tractor, trailer, car, and roadside streams give researchers a shared testbed for occlusion-heavy trucking.","key_machinery":"The load-bearing artifact is the TruckV2X dataset itself, generated by modeling a semi-trailer truck as an interactive agent in an urban driving simulator, with LiDAR and cameras distributed across the tractor and trailer and additional LiDAR/camera sensor suites on a connected car and roadside units. The analysis is carried by the occlusion recovery rate (ORR), defined as the fraction of an ego agent's occluded objects that cooperative agents can detect, which drives the paper's quantitative claims about trucks as both occlusion sources and perception enhancers. The benchmark uses early, intermediate, and late fusion configurations, including several intermediate-fusion baselines, to measure detection performance across agent roles.","core_discovery":"The paper's central claim is that TruckV2X is the first truck-centered cooperative perception dataset, providing synchronized multi-modal observations from tractors, trailers, connected cars, and roadside units. Its analysis shows that a tractor-trailer acts as both a severe occlusion source and a useful mobile perception platform: turning maneuvers sharply increase the objects a trailer hides from its own tractor, the truck blocks a substantial number of objects from other agents' views, and cooperative partners can recover a large fraction of those occluded objects. The benchmark results show that cooperative fusion, especially early fusion, raises mean average precision over ego-only baselines for truck, car, and roadside ego agents.","pith_inferences":["The paper does not test transfer to real trucks; a natural next step would be to pre-train perception models on this dataset and fine-tune on real truck sensor logs, checking whether the simulated occlusion patterns hold.","The occlusion recovery rate metric could be applied to other large vehicles, such as buses or construction equipment, whose blind zones follow similar size or articulation patterns; the paper itself does not make this extension.","The observed degradation of intermediate fusion with a roadside ego agent suggests future fusion architectures should be viewpoint-aware rather than car-centered, an issue the paper explicitly flags as needing further research."],"forward_implications":["A truck-centered cooperative perception benchmark now exists, letting methods be compared on heavy-vehicle occlusion scenarios rather than only light-vehicle data.","Trailer-mounted sensors measurably reduce the tractor's blind zones, supporting the paper's view of the truck as a mobile perception platform.","Cooperative fusion, particularly early fusion, improves detection over ego-only baselines for all ego agents in the simulated scenarios.","Occlusion recovery drops as inter-agent distance grows, suggesting proximity-aware agent selection is important for system design.","Intermediate fusion performs inconsistently when the roadside unit is the ego agent, indicating viewpoint differences among agents must be handled explicitly."],"supporting_citations":[{"why":"Represents the light-vehicle perception datasets whose lack of truck-specific scenarios motivates TruckV2X.","marker":"[6]"},{"why":"Supplies the cooperative perception dataset format and one of the intermediate-fusion baselines evaluated in the benchmark.","marker":"[8]"},{"why":"Provides the real-world vehicle-infrastructure cooperative perception baseline that TruckV2X extends to heavy vehicles.","marker":"[9]"},{"why":"Establishes the existing single-agent truck perception dataset that TruckV2X augments with trailer-mounted sensors and multi-agent cooperation.","marker":"[10]"},{"why":"Provides the urban driving simulator in which the semi-trailer truck model and all multi-agent scenarios were generated.","marker":"[11]"},{"why":"Serves as one of the intermediate-fusion cooperative perception baselines whose detection performance is benchmarked.","marker":"[21]"},{"why":"Serves as an intermediate-fusion baseline for joint perception in the benchmark experiments.","marker":"[22]"},{"why":"Contributes an intermediate-fusion baseline using cooperative attention over shared bird's-eye-view features.","marker":"[23]"},{"why":"Contributes a communication-efficient intermediate-fusion baseline included in the evaluation.","marker":"[24]"}],"fun_headline_variants":["First truck-centered cooperative perception dataset with multi-agent data","TruckV2X: truck, trailer, car, and roadside streams for occlusion-heavy scenarios","First truck-centric dataset to study blind spots and trailer occlusions","TruckV2X: multi-modal, multi-agent data for heavy-duty perception","Cooperative perception dataset centered on trucks tackles blind spots"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The dataset's value rests on the assumption that the simulated semi-trailer truck and its sensors reproduce real-world occlusion patterns closely enough that conclusions about cooperative perception transfer to physical trucks.","fun_headline_variants_meta":{"raw":{"variants":["First truck-centered cooperative perception dataset with multi-agent data","TruckV2X: truck, trailer, car, and roadside streams for occlusion-heavy scenarios","First truck-centric dataset to study blind spots and trailer occlusions","TruckV2X: multi-modal, multi-agent data for heavy-duty perception","Cooperative perception dataset centered on trucks tackles blind spots"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000791,"raw_usage":{"total_tokens":3446,"prompt_tokens":863,"completion_tokens":2583,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":479,"completion_tokens_details":{"reasoning_tokens":2490}},"tokens_in":479,"tokens_out":2583,"duration_ms":21119,"temperature":1.0,"reasoning_tokens":2490,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T17:53:17.842405+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"A real-world field test that records a tractor-trailer, a connected car, and a roadside unit at the same intersection configurations and compares the measured occlusion recovery rates and occluded-object counts with TruckV2X's; a systematic mismatch would show the simulated benchmark does not represent real truck occlusions.","supporting_citations":[{"cited_title":"nuscenes: A multimodal dataset for autonomous driving,","cited_arxiv_id":null,"evidence_quote":"Represents the light-vehicle perception datasets whose lack of truck-specific scenarios motivates TruckV2X."},{"cited_title":"Opv2v: An open benchmark dataset and fusion pipeline for perception with vehicle-to- vehicle communication,","cited_arxiv_id":null,"evidence_quote":"Supplies the cooperative perception dataset format and one of the intermediate-fusion baselines evaluated in the benchmark."},{"cited_title":"DAIR-V2X: A large-scale dataset for vehicle- infrastructure cooperative 3d object detection,","cited_arxiv_id":null,"evidence_quote":"Provides the real-world vehicle-infrastructure cooperative perception baseline that TruckV2X extends to heavy vehicles."},{"cited_title":"MAN TruckScenes: A multimodal dataset for autonomous trucking in diverse conditions,","cited_arxiv_id":null,"evidence_quote":"Establishes the existing single-agent truck perception dataset that TruckV2X augments with trailer-mounted sensors and multi-agent cooperation."},{"cited_title":"Carla: An open urban driving simulator,","cited_arxiv_id":null,"evidence_quote":"Provides the urban driving simulator in which the semi-trailer truck model and all multi-agent scenarios were generated."},{"cited_title":"F-cooper: Feature based cooperative perception for autonomous vehicle edge computing system using 3d point clouds,","cited_arxiv_id":null,"evidence_quote":"Serves as one of the intermediate-fusion cooperative perception baselines whose detection performance is benchmarked."},{"cited_title":"V2vnet: Vehicle-to-vehicle communication for joint percep- tion and prediction,","cited_arxiv_id":null,"evidence_quote":"Serves as an intermediate-fusion baseline for joint perception in the benchmark experiments."},{"cited_title":"Ermvp: Communication-efficient and collaboration-robust multi-vehicle percep- tion in challenging environments,","cited_arxiv_id":null,"evidence_quote":"Contributes a communication-efficient intermediate-fusion baseline included in the evaluation."}],"review_version":1}