DYNA-2 PROVES SCALING LAWS FOR ROBOTICS: 1 MILLION HOURS OF HUMAN VIDEO UNLOCKS ZERO-SHOT DEXTERITY

Summary from the TLDR newsletter

Dyna-2 is a world-action model (WAM) pre-trained on over one million hours of human video data. Its existence proves that scaling human video data predictably improves zero-shot performance on unseen robot hardware. Dyna-2 achieved an 87% pass rate in real-world zero-shot deployments at customer sites, vastly outperforming the 46% pass rate of its VLA predecessor, Dyna-1. The model uses a new one-step video generation distillation pipeline that drops inference latency by two orders of magnitude for downstream planning.