Dyna-2: Robot Task Improvement via 1M Hours of Human Video Pre-training

alifcoder · x · 2026-08-11

Dyna Robotics introduces Dyna-2, a world-action model pre-trained on over 1 million hours of human first-person video. The research reveals that this pre-training significantly improves robot task performance even without any robot trajectory data.

The study also uncovers several new scaling laws: world-action models exhibit scaling laws on human data across four orders of magnitude; this implies a scaling law on unseen robot data; and both data and objectives are crucial, with world modeling on video data being essential for cross-embodiment transfer.

Related event: Dyna-2: Pretraining on 1M Hours of Human Video Reveals Embodied Scaling Laws(15 posts)→

Original post →

More from Embodied

Embodied channel →