TLDR Tech / Hardware
Dyna-2 establishes a million-hour scaling law for robot learning
Dyna-2 is a video-diffusion world-action model pre-trained on more than 1 million hours of egocentric human-manipulation video, with no robot data included during pre-training. Dyna reports that prediction performance improved monotonically across a 1,000-to-1-million-hour training ladder, including zero-shot evaluation on 39 robot tasks from embodiments excluded from training. After limited robot post-training, normalized physical performance across 14 tasks rose from 20% to 53% as human-video pre-training scaled, with the largest model leading on nine tasks. Roughly 10 minutes of robot demonstrations were sufficient to post-train bottle-cap opening with two five-fingered robot hands.
Read the full story