Dyna-2.1: General Purpose Physical Agents with Hour-Long Autonomy

Last month, Dyna Robotics released the world-action model Dyna-2, calling it the first world-action model trained on over one million hours of data; it then released the physical robot agent Dyna-2.1, calling it the first physical agent to achieve reliable ultra-long-horizon whole-body autonomy. These two "firsts" form the main axis of the entire launch narrative: first a large model pretrained on human video, then an agent system built around it that can handle ultra-long real-world workflows.

But "first" is itself a judgment that only holds once an industry benchmark definition exists. The evidence pack explicitly notes that Dyna-2.1's "first to achieve reliable ultra-long-horizon whole-body autonomy" is a vendor self-positioning, lacking a comparable industry benchmark definition; whether Dyna-2 is truly the "first" world-action model trained on over one million hours of data also cannot be confirmed from a single source. In other words, what the publisher offers is positioning, not comparable evaluation results.

This distinction is not nitpicking. The core numbers of the entire report—the 99.4% success rate, 24 hours of continuous operation, 45-minute demonstration learning, the four-orders-of-magnitude scaling law—all come from Dyna's own statements in the interview or social posts, with no independent third-party verification. This evidence pack is also based solely on a single Latent Space podcast report, lacking Dyna's official technical report, paper, or independent evaluation as cross-confirmation. Therefore, this article's approach to these numbers is: preserve their values and context, but label them as source claims rather than confirmed facts.

Dyna says its model reached a 99.4% success rate on a continuous 24-hour restaurant napkin-folding task, and can recover after an entire stack of napkins is pulled down. This combination—a high success rate plus the ability to recover from its own failures—is precisely the most persuasive part of the "long-horizon autonomy" narrative: it tries to show that the robot can not only complete a single action, but handle disturbances over long runs.

However, the specific evaluation conditions, task definition, and statistical methodology behind the 99.4% success rate are all undisclosed, making reproducibility hard to judge. The 24 hours is a continuous runtime, but the source does not say whether there was human intervention during those 24 hours, whether the task was always the same napkin-folding action, or how failures were counted. These gaps make the number better understood as a demonstrative claim rather than a horizontally comparable performance metric.

The demo video also needs its context preserved. The Dyna-2.1 release came with an uncut demo video, which received about 3.12 million views on X. Uncut footage is more credible than a polished promotional video, but it is still material released by the vendor itself, and its representativeness cannot be independently verified. View counts indicate spread and buzz, not task success rates.