It also has exceptional robustness and zero-shot environment generalization; under various disturbances, the model continues to work. Dyna-2 is our first model to achieve an 87% quality and throughput rating during zero-shot deployment at novel customer sites. .

Beyond these scaling law results, Dyna-2 has many new capabilities that exceed our prior models. First, it has excellent language following capabilities through both scaling data and choosing the right world-modeling objective; it’s the first internal model that clears our internal language following benchmark.

The most important result in Dyna-2 is that the human data scaling law in fact implied a scaling law on never-seen robot data: scaling human data for pre-training can predictably improve zero-shot offline evaluation on robot data. In our blog, we also show that this scaling also transfers to on-robot evaluation via post-training on diverse robot embodiments.

Read our full technical blog for details: Dyna-2 established that there is no stopping in scaling: scaling human data up to millions of hours can predictably improve evaluation performance on human data. To ensure the emergence is metric agnostic, we demonstrated scaling laws on four distinct metrics.

Today we are introducing Dyna-2, a world-action model pre-trained on one million hours of human video. At this scale, for the first time, we discovered several new scaling laws: • world-action models exhibit scaling law on human data across four orders of magnitude, from 1000 to 1,000,000 hours, • this human data scaling law implied a scaling law on never seen robot data, • both data and objective matter; world modeling and scaling on video data are essential for cross-embodiment scaling transfer to emerge 🧵

Overall, we believe that Dyna-2 presents a new frontier in scaling embodied AI and charts an exciting direction forward to achieving general embodied intelligence. To find out more, check out our technical report : 1,000,000 hours is just a start. If you are interested in partnering with us on data or models, please reach out: [email protected]

We find that for the cross-embodiment scaling law to emerge, both training objective and data matter. World modeling and scaling on video data are essential.

Why did you strobe out the task space?

Why the hell is this video of a robot demo designed to caused photoepileptic seizures for no apparent reason?

Is this 1x speed?

The 87% at novel sites is the result that matters. Deployment still meets messy physical limits like sensor placement, latency, compute, heat, and actuator variance. Which hardware constraint caused the biggest drop from controlled tests to customer sites?

🚀 Looks promising! My DMs are open 📩

The flashing lights are so disturbing to watch...

Amazing concept. Would love to connect and discuss a collaboration 🚀@DynaRobotics