Teaching self-driving cars to see hidden danger

Self-driving cars must handle blind spots where they can't see other vehicles or pedestrians. This work models risk in occluded regions by combining traffic flow and collision hazards, then uses diffusion-based generation to create realistic training scenarios where occlusions matter. On Waymo's motion dataset, the method significantly outperforms existing occlusion-aware approaches, raising average time-to-collision margins by 67%.