Waymo Open Dataset

The Waymo Open Dataset is a large autonomous-driving dataset released by Waymo, captured by its self-driving fleet across multiple cities, times of day, and weather conditions. It is split into two main components: a Perception dataset of densely labeled sensor segments and a Motion dataset of mined driving scenarios and trajectories for behavior prediction, each maintained as an independent release.

What’s in it

The Perception data pairs high-resolution imagery from five cameras with five LiDAR sensors (one mid-range roof unit plus four short-range), all calibrated and time-synchronized, and includes millions of 2D and 3D bounding-box labels for vehicles, pedestrians, cyclists, and signs across roughly a thousand 20-second segments. The Motion dataset contains on the order of 100,000 scenes with agent tracks and high-definition map context for interaction modeling, and later releases add point-cloud semantic segmentation and keypoints.

Tasks, license & access

It supports 3D and 2D object detection, tracking, domain adaptation, occupancy and flow, and motion/behavior prediction, and it powers several public leaderboards and annual challenges. Successive versions have grown the segment count and added a modular, column-oriented (Parquet) format that lets users download only the sensors and labels they need. The data is free for research after account registration but is governed by the Waymo Dataset License Agreement, which restricts use to non-commercial purposes.

• Modality: multi-camera imagery plus multi-LiDAR point clouds, HD maps
• Scale: ~1,000+ perception segments; ~100K motion scenes
• Tasks: 3D/2D detection, tracking, motion prediction, occupancy flow
• License: Waymo Dataset License Agreement (non-commercial)
• Creator: Waymo LLC

Its sensor quality and scale make the Waymo Open Dataset a leading benchmark for perception and prediction in real-world autonomous driving.