“The fine-tuning for NVIDIA's Eagle Scale model used only 50 hours of motion capture data and 4 hours of teleoperation data, which constituted less than 0.1% of the total training mix.”