Dyna Robotics Dyna-2

2026-08-15 · Publisher : Dyna Robotics · Domain : Robotics

Dyna Robotics unveils Dyna-2, a world-action model pretrained on over one million hours of first-person human video, and claims the first human-to-robot transfer scaling law.

What changes

Pretrained on over one million hours of egocentric human video.
Data ladder tested from 1,000 to 1,000,000 hours.
Scaling law observed on never-seen robot data, per Dyna.
Head-to-head world-action model vs VLA, and zero-shot deployment.

### Background On August 15, 2026, Dyna Robotics released Dyna-2, a world-action model (WAM) pretrained on over one million hours of first-person human video. The company tackles an open question: which data source should be scaled so robots improve predictably? ### Specifications - **Data**: over one million hours of egocentric human video, with no robot data in pretraining. - **Scaling laws**: a ladder of models trained from 1,000 to 1,000,000 hours shows steady gains on held-out human data and, for the first time per Dyna, a transfer scaling law on robot data never seen during training. - **Real-world performance**: the study checks that this law translates into on-robot performance (results published by the company). - **Analyses**: an apples-to-apples WAM vs VLA comparison, zero-shot real-world deployment, instruction following through world modeling and one-step video generation. ### Why it matters Human video exists in far larger quantities than teleoperation data. If human-to-robot transfer follows a scaling law, video becomes a new growth axis for robot models without multiplying data-collection robots. ### Key takeaway Dyna Robotics moves from single-task reliability to the scaling race; Dyna-2.1 turned this model into a full physical agent at the end of September 2026.

Dyna Robotics Dyna-2

What changes

Links