Dyna-2

Dyna-2 is a world-action model (WAM) developed by Dyna Robotics, pretrained on over 1,000,000 hours of human video with no robot data in pretraining. It is documented in the August 2026 technical report 'Dyna-2: A 1-Million-Hour Scaling Law for World-Action Models'. The report demonstrates manipulation scaling laws from 1,000 to 1,000,000 hours of pretraining data, and shows the scaling law holds across the embodiment gap: held-out robot evaluation metrics improve monotonically as purely human pretraining data grows. The robot evaluation covers 39 held-out tasks on two bimanual YAM platforms (12 internal tasks and 27 from the xdof ABC dataset) with no adaptation. On real-robot deployment, Dyna-2 achieves an 87% pass rate under production criteria at customer sites, compared to 46% for Dyna-1 at an identical post-training budget, and generalizes significantly better when deployed zero-shot. Dyna-2 is also data-efficient in post-training: a bottle-cap untwisting task was post-trained on roughly 10 minutes of robot demonstrations. The report further documents robustness under extreme working conditions, instruction following, and one-step video generation, and includes a direct WAM-vs-VLA comparison.