TruckV2X Truck-Centered Cooperative Perception
Last updated: 2026-05-23
TruckV2X is a 2025 IEEE Robotics and Automation Letters dataset for truck-centered cooperative perception. It is useful because most V2X perception datasets focus on light vehicles or paired vehicle-infrastructure scenes, while long articulated trucks create different blind zones, trailer occlusions, and cooperator roles.
Related pages: V2X large-range and sequential datasets, infrastructure cooperative perception, collaborative fleet perception, SparseCoop, QuantV2X, V2X-Radar, autonomous trucking lane operations
Scope
| Item | TruckV2X coverage |
|---|---|
| Primary domain | Synthetic autonomous trucking cooperative perception |
| Source status | IEEE RA-L 2025 article, arXiv record, official project page, and Hugging Face dataset |
| Simulator | CARLA with Unreal Engine semi-trailer truck modeling |
| Agents | Tractor, trailer, connected autonomous vehicle, and roadside unit |
| Scale | 64 scenarios, 88,396 LiDAR frames, and 1.18M 3D bounding boxes on the project page |
| Core question | How tractor-trailer geometry changes occlusion, sensing, and cooperation benefit |
TruckV2X should be read as a benchmark and occlusion-analysis dataset, not as a cooperative perception model. It owns the truck-centered dataset gap; method pages such as SparseCoop, QuantV2X, CoSDH, and CoopTrack own the algorithmic fusion choices.
Sensors And Labels
| Agent | Sensors |
|---|---|
| Tractor and trailer | 64-channel LiDARs plus multi-camera RGB coverage; the project page lists 2 LiDARs and 5 cameras for tractor/trailer configurations. |
| CAV | 64-channel LiDAR plus four-camera coverage. |
| RSU | 64-channel LiDAR plus camera sensing from a fixed roadside viewpoint. |
Dataset records include synchronized point clouds, RGB images, per-agent metadata, calibration/pose information, and 3D object annotations. The public Hugging Face release is MIT licensed and includes train, validation, and test folders plus a dataset loading script; the dataset viewer currently cannot inspect it directly because the release uses a custom Python dataset script.
Labels support cooperative 3D object detection and truck-specific occlusion studies. The paper/project group evaluation around light vehicles, heavy vehicles, and vulnerable road users.
Truck-Specific Questions
| Question | Why it matters |
|---|---|
| Tractor self-perception | Long hood, high cab, trailer articulation, and large blind zones change what onboard sensors can see. |
| Trailer as cooperator | Trailer-mounted sensing can recover rear-quarter and side occlusions that tractor-only datasets miss. |
| Truck as occluder | Large trucks block CAV and RSU views differently from passenger vehicles. |
| Truck as mobile infrastructure | A truck can become a useful remote sensing platform for nearby road users. |
| Articulation angle | Trailer pivot changes occlusion and sensor overlap during turns and yard-style maneuvers. |
The project page reports that tractor-trailer combinations create substantially more occluded area than passenger cars within short range and that trailer pivot can worsen blind zones during large turns. Treat these as dataset-analysis findings for the TruckV2X simulator, then re-check them on target fleets or real yards before using them as operational evidence.
Tasks And Metrics
| Task | Metric or output |
|---|---|
| Cooperative 3D object detection | AP and mAP at IoU 0.3, 0.5, and 0.7 |
| Ego-role comparison | Truck, CAV, RSU, tractor, and trailer cooperation modes |
| Occlusion recovery | Occlusion Recovery Rate for objects that an ego agent misses but cooperators observe |
| Class-sliced detection | Light vehicle, heavy vehicle, and VRU performance |
| Fusion comparison | Early, intermediate, late, and no-fusion baselines |
The project page reports benchmark experiments across eight cooperative perception methods. It highlights early fusion as strongest for truck ego at IoU 0.5 and notes that vulnerable-road-user detection remains difficult because small objects generate sparse LiDAR evidence.
Best Use
Use TruckV2X to:
- test whether cooperative perception methods handle articulated heavy vehicles rather than only passenger-car scenes;
- benchmark truck, trailer, CAV, and RSU collaboration modes separately;
- measure occlusion recovery around a truck instead of only global AP;
- prototype data schemas for trailer-mounted sensors and truck-as-cooperator messages;
- compare against general V2X datasets before collecting real yard, port, highway, or airside data.
It is especially useful when a project has large, articulated, or high-occlusion vehicles but lacks a target-domain dataset.
Domain Fit
| Domain | Fit | Notes |
|---|---|---|
| Road AV / trucking | Strong research fit | Directly targets autonomous trucking and heavy-vehicle cooperative perception, but it is synthetic. |
| Airside AV | Conditional proxy | Long GSE, baggage trains, catering trucks, fuel trucks, and fixed infrastructure create analogous occlusion patterns; aircraft geometry, GSE classes, FOD, wet aprons, jet blast, and ramp rules are missing. |
| Logistics yards and ports | Moderate to strong proxy | Articulated and heavy equipment plus fixed infrastructure transfer conceptually, but local layouts and object taxonomies differ. |
| Mining and construction | Conditional | Large-machine occlusion transfers, while terrain, dust, slopes, buckets, workers, and site geometry need local data. |
| Warehouses and campuses | Limited | Cooperative-agent ideas transfer, but road-scale trucks, CARLA scenes, and sensor geometry are less representative. |
For airside or industrial transfer, do not treat TruckV2X as final evidence. Use it to design sensor placement, cooperator roles, and occlusion metrics before collecting target-domain logs.
Limitations
- It is synthetic CARLA/Unreal data, not real-world trucking logs.
- It does not include airport-specific aircraft, GSE, stand markings, jet bridges, FOD, chocks, cones, hoses, glycol film, or wet-apron multipath.
- It does not by itself validate V2X bandwidth, latency, packet loss, trust, cybersecurity, or time synchronization.
- Dataset licensing and generated-asset constraints should be reviewed before commercial reuse.
- AP/mAP can hide rare safety-critical failures; report occlusion, class, ego-role, and distance slices.
- Trailer-sensor value depends on mounting, calibration, wiring, maintenance, and operational coupling that may differ from real fleets.
Implementation Notes
- Keep tractor, trailer, CAV, and RSU as separate agents in data manifests.
- Preserve articulation angle, relative pose, timestamp, calibration, and agent role fields.
- Report truck ego, tractor ego, CAV ego, and RSU ego scores separately before averaging.
- Add latency and packet-drop replay if using TruckV2X to evaluate deployable V2X methods.
- For trucking operations, pair the dataset with lane-operation evidence, terminal handoff rules, inspection logs, and weather ODD gates.
- For airside or yard transfer, add classes for local equipment and use occlusion-zone metrics around large vehicles, containers, aircraft, or fixed structures.
Sources
- TruckV2X official project page: https://xietenghu1.github.io/TruckV2X/
- TruckV2X Hugging Face dataset: https://huggingface.co/datasets/XieTenghu1/TruckV2X
- TruckV2X arXiv record: https://arxiv.org/abs/2507.09505
- TruckV2X arXiv PDF: https://arxiv.org/pdf/2507.09505
- IEEE DOI: https://doi.org/10.1109/LRA.2025.3592884