Twenty thousand hours of hands
Two and a quarter robot-years of continuous dual-arm operation, recorded on purpose. That is not a dataset. That is a factory built to produce one.
What that number costs
Twenty thousand hours is roughly two and a quarter years of continuous operation, and it is real rather than simulated. It is a capital expenditure disguised as a dataset, available only to an organisation willing to run robot farms for the express purpose of generating training data.
That is a moat of a specific and uncomfortable kind. It cannot be replicated by a better idea. It can only be replicated by the same expenditure, which means the set of organisations able to compete at this layer is decided by balance sheet rather than by insight.
Nine configurations is the harder achievement
Training across multiple embodiments is what separates a policy that works on one arm from a foundation model. It is also where most cross-embodiment work fails, because the mapping between different kinematics is not free — a motion that is natural on one arm may be unreachable on another, and the model has to learn the correspondence rather than assume it.
The alternative approach is learning latent actions from video with no action labels at all, which trades data cost for the risk that the latent space does not transfer. Both are live. Neither has won.
The infrastructure is catching up
A serving system for robotics foundation models in robot factories is the unglamorous end of this — versioning, latency guarantees, failure handling across many machines.
The requirements are harder than language serving in one specific way. A language serving system can queue a request. A policy driving a physical arm cannot, because the world does not pause while inference completes, and a late action is frequently worse than no action.
Related work on thermal-perceptive models for safety-critical manipulation and on gaze-conditioned policies points the same direction: the open questions are increasingly about deployment conditions rather than capability ceilings. That is what a field looks like shortly before it becomes an industry.
arXiv — ROSA: a robotics foundation model serving system for robot factories → · arXiv — Safe-Night VLA: thermal-perceptive vision-language-action models for safety-critical manipulation → · arXiv — Nautilus: from one prompt to plug-and-play robot learning →