D2E: Scaling Vision-Action Pretraining on Desktop Data for Transfer to Embodied AI Paper β’ 2510.05684 β’ Published Oct 7, 2025 β’ 141