Waymo Open Dataset
Waymo
Lidar, camera, and motion logs from real robotaxis, released for anyone doing research.
Sensor data and labels from Waymo's self-driving vehicles in several U.S. cities, in three parts: Perception (synchronized lidar and camera with 3D boxes), Motion (object tracks for predicting what road users do next), and End-to-End driving. It is one of the largest public autonomous-driving data sets and backs annual research challenges.
Access
- Cost
- FreeFree after signing in with a Google account and accepting the license.
- License
- Open, non-commercial onlyWaymo Dataset License Agreement for Non-Commercial Use
- Sign-up
- Free account
- How you get it
- Bulk downloadTFRecord, Parquet
What's in it
- One row is
- One 20-second driving segment (Perception) or scenario (Motion) with labeled objects
- Covers
- U.S. cities where Waymo operates, including Phoenix and San Francisco
- Years
- 2019-present
- Updated
- Annual
- Size
- Perception has about 2,000 segments; Motion has over 100,000 scenarios
Quality
High-quality sensors and labels with versioned releases and benchmarks; commercial use is barred.
- Non-commercial license
- Few cities and mostly good weather
- Faces and license plates are blurred
Connects to
- Kinetics-700: Video understanding benchmark for general human actions, complementing driving-specific video
- COCO (Common Objects in Context): 2D detection pretraining before fine-tuning on driving imagery
Questions it could answer
- How often do pedestrians cross outside crosswalks in the Motion scenarios?
- How do detection errors grow with distance from the vehicle?