Learning a Unified Risk Map for Autonomous Driving in Partially Observable Environments

Researchers introduced a unified risk map for autonomous driving that combines traffic-flow risk and collision risk in partially observable scenes.

They also used diffusion-based scenario generation to create realistic occluded interactions for training and evaluation.

On the Waymo Open Motion Dataset, the approach outperformed prior occlusion-aware baselines.

The framework could help planners reason about hidden hazards more safely and consistently.