Deep learning models for optical flow estimation, like RAFT, achieve high accuracy but often lack robustness under real-world conditions. In this thesis, we investigate whether data augmentation can improve robustness against dataset shifts, image corruptions and adversarial attacks. To this end, we integrate four augmentations (Gaussian noise, Gaussian blur, JPEG compression and snow) into the RAFT training pipeline with varying severity levels and augmentation proportions. Additionally, we consider two training approaches: pretraining only and pretraining with additional fine-tuning, where augmentations are applied either during pretraining or during fine-tuning. Models are evaluated on Sintel and KITTI using three robustness metrics. Our experiments show that data augmentation can enhance robustness, but its effectiveness depends strongly on augmentation type, severity and proportion as well as the training setup. Gaussian noise particularly improves robustness to corruption and adversarial attacks but at the cost of accuracy. Gaussian blur preserves accuracy yet offers limited robustness gains, while JPEG compression and snow provide more balanced effects. Overall, mild augmentations yield the most favorable trade-offs, though a clear accuracy-robustness trade-off remains.