What happened?

Dyna Robotics has released Dyna-2, a world-action model for robot manipulation. It was pre-trained on more than one million hours of egocentric human video — roughly 170 years of continuous waking experience.

The question is clear: robot learning has long been bottlenecked by action-labelled data, and that data can only be produced through teleoperation, which means producing it deliberately. Dyna-2 tests whether ordinary human video can substitute.

What was measured

The research team trained a data ladder from 1,000 to 1,000,000 hours and measured what scales. Three results are reported:

  • A scaling law: a scaling law was observed on human data.
  • Transfer: the first transfer of that law to unseen robot data was demonstrated.
  • Mechanism: evidence was offered that video prediction drives the transfer.

Architecture

Dyna-2 is a world-action model: a single generative model that denoises future video and a future action chunk, jointly or separately, on a video-diffusion backbone.

Architecturally it is a mixture of transformers. Video and action are tokenised separately and get distinct DiT layer stacks that attend to each other. Proprioception feeds directly into the action transformer. Video tokens use causal masking; action tokens use bidirectional self-attention and attend to context video tokens. Video tokens cross-attend to text, but text does not directly influence the action.

How it is used

The model ships as a vendor-operated system rather than as downloadable weights. Dyna Robotics has announced no public checkpoint, API or licence; deployment at this stage means buying a Dyna robot cell, not self-hosting a model.

Per the company's 10 August 2026 announcement, Dyna-1 robots already run in production in hotels, restaurants and laundromats. That points at the target market: mid-market service operators and multi-site enterprises with repetitive, stationary manipulation work. It is not a fit for solo builders or research labs wanting local inference.

Which jobs

The 14 post-training tasks map cleanly onto real work: trash tray clearing, first-aid kitting, tote construction, food scooping, rope tying, hanger preparation and targeted drink retrieval from a fridge. The industries listed are hospitality, commercial laundry, food service, light assembly and kitting, and facilities cleaning.

Why the question matters

The most expensive input in robot learning is data collection. Teaching a robot a task requires a human to drive that robot by remote control many times over, with every attempt recorded and labelled. The process is measured in hours and scaling it depends directly on human labour.

Human video, by contrast, already exists. A million hours of egocentric recording can be gathered without anyone having filmed it to train a robot. That is the economic weight of the question: if ordinary video can stand in for action-labelled data, the cost curve of robot learning changes entirely.

What is not settled

The scaling law and transfer findings rest on the company's own work; because the weights cannot be downloaded, independent verification is not possible for now. That is the basic difference from models released with open weights in the same period: the numbers are not reproducible.