“The Groove2Dreams model was created by fine-tuning a video world model, pre-trained on hundreds of millions of hours of YouTube videos, with a small amount of in-house teleoperation data.”