New preprint: RAP. Lightweight 3D rasterized views can augment imitation learning data for end-to-end driving.