NVIDIA presented GR00T 2 at the Conference on Robot Learning, according to Interesting Engineering (Nov 9). It’s described as a world action model that predicts future video and robot motion together, and it learns from human videos, handheld grippers and teleoperated robots. The training mix NVIDIA states: 402,730 hours of first person human video, 23,500 hours of robot teleoperation and 10,184 hours of handheld gripper demos. By my math, human video is over 90% of the hours listed. The results are NVIDIA’s own: 40% average success on 50 unseen tasks, rising to 95% after five demos per task and a 30minute fine tune. That 95% is an average, not a per task rate. My take: the biggest pile of training data in this announcement was captured by people, not robots. That makes who captured it, and whether they were credited and paid, a real question. It’s the part I follow at @vangrid_io: phone capture, USDC bounties settled on acceptance, fingerprints anchored on Base. Would you contribute your own footage to train robots?