The Waymo Open Motion Dataset is a component of the broader Waymo Open Dataset released by Waymo to support research in autonomous driving, machine perception, and planning. It contains over 100,000 driving scenes, each capturing 20 seconds at 10 Hz, spanning more than 570 hours of recorded data across roughly 1,750 kilometers of roadways.
Each scene includes richly annotated trajectories of various road agents (vehicles, pedestrians, cyclists), along with highly accurate 3D bounding boxes and HD maps, enabling development of models that predict the future paths of multiple interacting agents in a shared space.
This dataset is especially useful for building and evaluating motion-forecasting systems that account for interactive dynamics rather than isolated agent prediction. It also serves as a benchmark in the autonomous driving community and fuels advances in joint agent motion prediction, behavior modeling, and planning
Use case:
Multi-Agent Motion Forecasting and Behavior Modeling for Autonomous Driving
| Example Observation | Ego: speed (m/s), acceleration (m/s²), heading (rad), position x, y (m). Road context: polyline of the current lane (10 points), curvature angle, local speed limit. Neighboring agents (k ≤ 15): relative position, speed, acceleration, and type (car, pedestrian, cyclist). Traffic signals: light state (red, orange, green), remaining time. Timestamp: time within the traffic light cycle, time of day. |
| Example Action | at∈{−2m/s² (brake),0m/s² (maintain),+2m/s² (accelerate)} They are reconstructed by discretizing the real acceleration between t and t + 1 extracted from the Waymo traces. |
| Example Reward | +0.1 for progress (m traveled in the target lane) –1.0 for collision or crossing a solid line –0.3 for emergency braking (>4 m/s²) –0.05 per second spent stopped while the light is green |