“NVIDIA's Eagle Scale model was pre-trained on 21,000 hours of in-the-wild egocentric human video data with no robot data included.”