Our method introduces a fully pose-free feedforward framework DGGT for reconstructing dynamic driving scenes directly from unposed RGB images. The model predicts camera poses, 3D Gaussian maps, ...