AevaScenes

What It Is

AevaScenes is an open-access FMCW 4D LiDAR and camera dataset for autonomous vehicle research.
Aeva announced the dataset on September 30, 2025.
It features synchronized FMCW LiDAR and RGB camera data.
It includes object velocity measurements, semantic segmentation, tracking, and lane-line annotations.
The dataset is available for academic and non-commercial use through the official AevaScenes site.
It is a dataset, not a model architecture.

Provide public data from FMCW LiDAR, where each point can carry range and radial velocity information.
Expose motion information directly at sensor measurement time rather than deriving all velocity from multi-frame tracking.
Pair 4D LiDAR with cameras so researchers can study detection, segmentation, tracking, motion forecasting, scene flow, and calibration.
Include ultra-long-range annotations up to 400 m.
Make FMCW-specific perception research possible without private sensor access.
Show how velocity-per-point changes the design space for object detection and prediction.

Dataset input: FMCW 4D LiDAR point clouds.
Dataset input: synchronized high-resolution RGB camera images.
Dataset input: calibration and sensor metadata.
Labels: object detection, semantic segmentation, tracking, and lane-line annotations.
Format: PCD point clouds, JPEG images, and JSON annotations according to the announcement.
User output: trained or evaluated perception models for FMCW LiDAR and camera fusion.

The release describes 100 curated sequences.
It contains 10,000 frames at 10 Hz.
Sensor suite: 6 Aeva FMCW LiDAR sensors.
Camera suite: 6 high-resolution RGB cameras matched to wide and narrow field-of-view sensing.
The announcement reports about 200 GB total, roughly 2 GB per sequence.
Data was captured using Aeva Mercedes Metris test vehicles in and around the San Francisco Bay Area.

Aeva positions the dataset for object detection, semantic segmentation, tracking, motion forecasting, scene flow, and trajectory estimation.
The release includes 50 percent highway and 50 percent urban sequences.
It includes 50 percent day and 50 percent night sequences.
All sequences in the announcement are clear weather with dry road surfaces.
Evaluation protocols should isolate the value of per-point velocity versus position-only LiDAR.
The dataset is new enough that independent benchmark leaderboards may still be immature.

Public FMCW LiDAR data is rare, making this a high-value sensor research resource.
Per-point velocity can reduce the latency of motion detection and track initialization.
Long-range annotations up to 400 m support high-speed and early-warning research.
Multi-camera pairing supports fusion and cross-modal calibration studies.
Day/night balance helps evaluate lighting robustness.
Tracking and segmentation labels make it more useful than box-only releases.

Clear-weather, dry-road collection does not cover rain, fog, snow, de-icing spray, or jet blast.
Road scenes do not include aircraft, ramps, cones, dollies, fuel trucks, or ground crew workflows.
FMCW velocity is radial; lateral motion still needs geometry, tracking, or fusion.
Dataset license is non-commercial, which can restrict production training use.
Sensor configuration is Aeva-specific and may not transfer perfectly to other FMCW LiDAR vendors.
Long-range road annotations do not guarantee close-range aircraft-clearance accuracy.

Very relevant as the first practical public dataset for FMCW LiDAR perception design.
Per-point velocity is valuable for detecting moving GSE, personnel, jet blast particles, and approaching vehicles with lower latency.
Clear-weather limitation means it cannot validate the most important airside weather claims.
Airport transfer requires new data around aircraft metal surfaces, wet aprons, night lighting, and de-icing operations.
The 6-LiDAR setup is conceptually close to multi-LiDAR airside vehicles, though sensor placement will differ.
Best use is pretraining and architecture prototyping before collecting proprietary airside FMCW data.

Extend point-cloud schemas to preserve velocity fields, not just XYZ intensity.
Update ROS PointCloud2 messages with an optional radial_velocity field for FMCW-aware pipelines.
Benchmark single-frame velocity-based detection against multi-frame ToF LiDAR tracking.
Evaluate how much per-point velocity improves track birth, stop/start detection, and prediction horizon.
Keep a compatibility path that drops velocity for existing LiDAR models.
Combine with radar in airside tests because radar and FMCW LiDAR have complementary weather and reflectivity behavior.