Motion estimation from MRI is important for image-guided radiotherapy. Specifically, for online adaptive MR-guided radiotherapy, the motion fields need to be available with high temporal resolution and a low latency. To achieve the required speed, MR acquisition is generally heavily accelerated, which results in image artifacts. Previously we have presented a deep learning method for real-time motion estimation in 2D that is able to resolve image artifacts. Here, we extend this method to 3D by training on prospectively undersampled respiratory-resolved data showing that our method produces high-quality motion fields at R=30 and even generalizes to CT without retraining.