DYNA-2

Dyna Robotics' 2026 world action model, pretrained on a million hours of first-person human video.

DYNA-2 is Dyna Robotics' second-generation model, released in August 2026, a world action model (one that jointly predicts future video frames and future actions). It uses a video diffusion model as its backbone, with a hybrid Transformer that processes the video stream and the action stream separately, trained with flow matching so it can denoise future video and actions either jointly or separately. Its pretraining data is more than a million hours of head-mounted, first-person human video, covering everyday activities like cooking, tidying up, folding clothes, and assembly. The company reports observing a scaling law from human data to robot performance: as human pretraining data increases in stages from 1,000 hours to 1 million hours, the average normalized performance (relative to the achievable ceiling) across 14 real-robot tasks rises in step, to 20%, 28%, 45%, and 53%. In the company's layered system, DYNA-2 serves as System 1, taking direction from DYNA-VLM, which handles reasoning.

ExampleAccording to the company, when folding napkins, DYNA-1 manages about 35 per hour at a 75% acceptable-quality rate, while DYNA-2 reaches 95 per hour at 93%; customer Din Tai Fung is now expanding Dyna's robots from a pilot restaurant to more of its locations and central kitchens.

Also called

Dyna-2

Related

DYNA-1 · World Action Model · Egocentric Video · Pretraining on Human Videos · Scaling Law · Dyna Robotics

Sources

Dyna-2: A 1-Million-Hour Scaling Law for World-Action Models (Dyna Robotics, 2026-08)
Not Just a Model, But a Product (Dyna Robotics, 2026-08)
Dyna Robotics 官网 (Chinese)

As of

2026-08