A Dataset and Benchmark for FMCW LiDAR Perception
AevaScenes is a comprehensive multi-modal dataset designed to advance research in FMCW (Frequency-Modulated Continuous Wave) LiDAR perception. The dataset features synchronized data from 6 FMCW LiDARs and 6 high-resolution cameras mounted on a vehicle, captured across diverse urban and highway environments in the San Francisco Bay Area.
AevaScenes features the longest-range FMCW LiDAR data ever released to the public, delivering over 400 meters of range with per-point velocity—enabling researchers and developers to explore perception capabilities beyond the limits of traditional datasets.
For the first time, FMCW LiDAR enables detections up to 400 meters by measuring instantaneous radial velocity per point directly at the sensor. Aeva's Doppler-based approach enhances long-range perception, enabling earlier detection of moving objects and improving overall safety for autonomous navigation.
AevaScenes provides annotations—including 3D bounding boxes, lane lines, and semantic segmentation—at ranges up to 400 meters. This unprecedented depth of annotation empowers research in long-range perception, planning, and tracking beyond the limits of existing datasets.
Unlike traditional LiDARs, FMCW LiDAR delivers high dynamic range with negligible blooming around retroreflective surfaces such as road signs, botts-dots and license plates, resulting in sharper object boundaries and more accurate perception, even in challenging high-reflectivity scenarios.
A small subset of the highway/city/day/night sequences are available to see using the web visualizer here AevaScenes Web Visualizer.
- Visit scenes.aeva.com/downloads.
- Register and agree to the license terms.
- We will email you the signed_urls.txt file which contains the download links.
- Download the dataset using the provided download script and signed_urls.txt:
# Download the dataset (after signing up and obtaining access)
mkdir -p data/aevascenes_v0.1
bash scripts/download_dataset.sh --url-file signed_urls.txt --output data/aevascenes_v0.1
# Extract the dataset
cd data/aevascenes_v0.1
for f in *.tar.gz; do tar -xvf "$f"; donePlease see Dataset.md for details about the dataset and schema.
Please see Getting Started.md to get started with using the Python SDK. Here's a quick overview.
# Visualize a single sequence
python examples/visualize_aevascenes.py --dataroot <DATA_ROOT> --viz-mode sequence --sequence-uuid <UUID> --color-mode [velocity/reflectivity/semantic]
# Visualize a single sequence with points projected
python examples/visualize_aevascenes.py --dataroot <DATA_ROOT> --viz-mode sequence --sequence-uuid <UUID> --color-mode [velocity/reflectivity/semantic] --project-points
# Visualize random sampled frames from all sequences
python examples/visualize_aevascenes.py --dataroot <DATA_ROOT> --viz-mode sampled --color-mode [velocity/reflectivity/semantic] --project-pointsThe AevaScenes dataset is provided under the AevaScenes Dataset License Agreement for non-commercial use only. The AevaScenes Python SDK is licensed under the MIT License.
- Dataset License: AevaScenes Dataset License
- Code License: MIT License
If you use AevaScenes in your research, please cite our work using the following BibTeX.
@misc{aevascenes,
title = {AevaScenes: A Dataset and Benchmark for FMCW LiDAR Perception},
author = {Narasimhan, Gautham Narayan and Vhavle, Heethesh and Vishvanatha, Kumar Bhargav and Reuther, James},
year = {2025},
url = {https://scenes.aeva.com/},
}We welcome contributions to improve the AevaScenes dataset and toolkit! Please see our contributing guidelines and submit pull requests for:
- Requests for data diversity.
- Bug fixes and performance improvements.
- New visualization features.
- Dataset utilities and analysis tools.
- Website: scenes.aeva.com
- Issues: GitHub Issues
- Email: research@aeva.ai