The first million hours showed that human experience can scale physical intelligence. The next frontier is to scale not only how many hours we collect, but how much learnable physical information each hour contains.
For years, one of the central questions in robotics has been whether Physical AI can follow a scaling path similar to language and vision. Large language models benefited from an abundant substrate: text already existed at internet scale, and vision models inherited billions of images. Robotics never got that. High-quality robot demonstrations are expensive to produce, tied to specific embodiments, and limited by the rate at which physical robots can operate. This has made data—not only models or compute—one of the defining bottlenecks of Physical AI.
Dyna Robotics' Dyna-2 is an important data point for how that bottleneck breaks. Pre-trained on more than a million hours of human egocentric video, it improves monotonically with human experience across four orders of magnitude, and that improvement survives the embodiment gap: models pre-trained on more human data did better on held-out robot data they had never seen. So is Genesis AI's GENE-26.5, but they break it in a different direction.
The question is no longer whether human experience can scale physical intelligence, but what to scale after the first million hours. My view is that the next generation of Physical AI data will scale along two distinct but coupled axes: Experience Scale is how many hours of physical experience you have; Experience Density is how much learnable physical information each hour of experience contains. Both matter, and each challenges the physical intelligence field in different ways.
Dyna-2 asks a simple question: how far can physical intelligence scale when we add more human experience?
Its answer is encouraging. Across nested datasets of 1K, 10K, 100K, and 1M hours, prediction performance improves monotonically with pre-training scale. More consequentially, zero-shot testing on robot data suggests that sufficiently broad human experience can begin to produce cross-embodiment transfer through scale itself.(technical report)
The subtler result is that more action labels alone are not enough. In Dyna's ablations, action-only training does not show the same cross-embodiment scaling behavior. Adding video prediction, and especially co-training on additional human video, made cross-embodiment generalization improve with scale.
The lesson is bigger than the one-million-hour number. Human experience appears to remain a useful scaling substrate well into foundation-model territory — which gives us the first axis, Experience Scale. But its value comes from learning how the physical world evolves under interaction, not simply from accumulating more labeled actions, and that is already a hint of the second.
GENE-26.5 asks the complementary question: how much more useful can human experience become when we preserve more of its physical structure?
Genesis frames today's data problem as a trade-off between scale and fidelity. Third-person and egocentric video scale well but leave much of an interaction implicit. More instrumented collection can expose hand state, contact, and tactile information, but historically at higher collection cost. GENE's data engine combines these sources rather than choosing one: third-person video for breadth, egocentric video for natural interaction, and instrumented gloves for high-fidelity hand motion and tactile signals.
Its manipulation framework also makes clear what richer physical experience contains. Genesis evaluates tasks through spatial precision, temporal composition, contact richness, contact coordination, and tool-mediated interaction—properties that describe not just what task happened, but how physical interaction unfolded in space, time, and contact.
The lesson is that an hour of human activity becomes more valuable when the data preserves more of the physical state that generated it — a point Genesis makes in their own framing, that the constraint is not volume alone but whether the capture retains the richness of the interaction. The commercial payoff is downstream: Genesis reports many difficult skills reaching autonomous execution on under an hour of task-specific robot data. Denser human experience is what buys down the expensive robot-specific data.
That gives us the second axis: Experience Density.
The obvious continuation is 1M → 10M hours and beyond, and the evidence so far says that continuation pays: none of the published curves has flattened, and larger collections of human experience are on track to become the pre-training substrate for more general robotic models. But the transition is not a bigger version of the same collection project. It is a different infrastructure regime.
Text pre-training inherited its substrate. Physical AI has to manufacture it: people perform tasks, sensors capture them, hardware stays calibrated, and the recording is then synchronized, validated, annotated, and converted into something a training system can consume.
So the scaling law is also a manufacturing law, and that changes which metric the business runs on. Cost per recorded hour is easy to measure and is what procurement asks for. What governs the business is cost per usable training hour — one that cleared calibration, sync, annotation, and QC for the supervision actually being bought. What governs the model is one step further out: cost per unit of new physical information — the measure that tells you whether each additional decade of hours is still buying capability.
That gap widens with scale, because the pipeline is a series of gates and yield is multiplicative. Four gates at 95% give 81% end-to-end; push every gate to 99% and yield reaches 96%. At ten-million-hour volumes those fifteen points are the difference between a corpus that gets delivered and one that gets reprocessed, which is why the next order of magnitude is as much a manufacturing achievement as a research one. Adaptive scheduling, QC close to capture, verification concentrated on ambiguous samples: this is the industrial capability that makes 10M hours deliverable at a price the field can absorb.
At small scale, volume and diversity rise together: 1,000 to 10,000 hours naturally adds people, homes, objects, viewpoints. Past a certain point they stop moving in lockstep. The ten-millionth hour resembles the corpus more closely than the ten-thousandth did — unless somebody put it there on purpose.
That is an argument for collecting more, and for collecting with intent. A corpus that reaches ten million hours and holds its slope is one where acquisition is steered by what is already covered: which environments, objects, and configurations are thin, and which recordings would fill them. Coverage-aware acquisition is what keeps the ten-millionth hour worth as much to a model as the ten-thousandth was, and it is the capability that separates a collection operation from a recording service.
So the frontier question is not only how many hours were added but how much new physical experience they contributed — the marginal information gain of the next sample against the corpus already held. Which brings us to the second axis.
If Experience Scale measures how much experience we have, Experience Density measures how much useful and learnable physical information sits inside each unit of it.
Two one-hour recordings make the point. One contains repetitive tabletop pick-and-place under fixed conditions. The other contains bimanual manipulation across changing states, with precise geometry, varied contacts, coordinated fingers, and tool use. Equal duration, nowhere near equal learning value. At ten thousand hours, that difference is noise; at ten million it is the budget.
One clarification, since this is where the framework is easiest to attack: Axis I coverage is breadth — more environments, people, and objects across the same kind of recording. Axis II is depth — how much of each interaction the recording preserves. Adding a thousand new kitchens is Axis I; capturing contact and object states inside the kitchens you already have is Axis II.
The first generation of large embodied datasets counted tasks, scenes, objects, participants. Those metrics remain useful, but task diversity is only a proxy for physical state-space diversity.
Take one instruction — open the drawer. A task-centric dataset counts one task. A state-centric view sees many distinct situations: drawer closed, partially open, obstructed; handle grasped off-axis; the other hand occupied; the drawer unusually heavy; contents shifting against the motion. These occupy different regions of state-action space, and a policy that has seen only the first cannot open drawers.
But state coverage should go one step further. For manipulation, contact is part of the state.
Two executions can begin and end in nearly identical object configurations while relying on very different physical interactions: one finger pushing versus a full-hand grasp; one stable contact versus several coordinated contacts; direct manipulation versus using a tool; slow quasi-static motion versus a precisely timed interaction.
GENE-26.5's evaluation axes make this distinction explicit: spatial precision, timing, contact richness, contact coordination, and tool-mediated interaction are separate dimensions of manipulation capability.
So the next generation of coverage metrics should ask not only which tasks and object states are represented, but which interaction states are represented.
Dyna-2 derives its pseudo-actions from 3D hand tracks alone, and that minimal state demonstrably worked. The question is what the next increment buys.
A denser representation would expose
— hand pose, body configuration, object pose and geometry, contact, and surrounding 3D environment — with the training signal being the transition rather than the observation at time . A world model fits that transition whether or not the dataset represents it explicitly, and every element of left unrecorded is one the model must infer from pixels: sometimes fine, sometimes the difference between learning contact physics and memorizing appearance. The fundamental unit of Physical AI data is not the frame or the clip but the physical interaction and the state transition it produces.
The elements of do not cost the same, though, and an honest roadmap prices them: hand pose is solved and cheap, body and environment geometry are moderate with stereo or SLAM, object pose is expensive and needs instrumented environments or strong priors, and contact is hardest to measure while richest per bit. My ordering is body and environment first, object pose where the task family justifies it, contact last — the simple version before the complete one.
Observing human activity and making it useful to arbitrary robots are different problems: a human wrist trajectory is not a robot joint trajectory, and a five-finger grasp is not a parallel-jaw command.
But much of the underlying interaction is representable independently of the actor: where contact occurs, how an object moves after contact, which grasp stabilizes it, how a hinge constrains motion. The richer those signals, the easier it is to turn human experience into supervision that transfers. Call that embodiment transferability; it is the practical target of this axis.
Dyna-2 bridges part of that gap with no human-robot alignment at all, so take it at face value: scale alone was sufficient. Whether it is efficient is the open question, and efficiency decides whether the next order of magnitude gets funded. Density should not mean annotating for its own sake — it should mean preserving the information that survives a change of embodiment.
The two axes are not competing philosophies, and neither substitutes for the other. No amount of annotation on a small dataset buys the breadth of millions of hours of real activity, and ten million hours return their full value only when each one is made to carry more.
| Low Experience Density | High Experience Density | |
|---|---|---|
| Low Experience Scale | Narrow demonstrations | Rich but limited research datasets |
| High Experience Scale | Massive raw experience | Frontier Physical AI Data |
A useful way to hold them together is an accounting identity, not a fitted law:
It is a bookkeeping frame for where the next dollar goes — a decision every data team here is already making implicitly. Scale expands the world the model has met; Density increases how much it learns from each encounter.
If the industry builds the next order of magnitude, the result should not be a larger video warehouse. A 10M-hour frontier corpus should behave more like a physical experience engine, and four of its properties are concrete enough to build against and to check.
It should know what it already holds, at the level of state coverage rather than hours and task labels. It should estimate where that coverage is thin. It should acquire against that estimate alongside its volume targets, seeking new states as well as new hours — rare interactions, difficult geometries, underrepresented object classes. And it should route each hour to the level of supervision its expected value justifies, since full geometric annotation is worth paying for on some recordings and wasted on others.
That changes the central KPI of embodied-data infrastructure. The question moves from how many hours you can collect toward how fast you can increase effective physical experience per dollar. Three numbers make that measurable: end-to-end pipeline yield, marginal information gain per delivered hour, and coverage against a declared state taxonomy. Only the first is well-defined today. The second and third are open problems, and whoever defines them first will set the standard the market prices against.
Dyna-2 and GENE-26.5 are not the only signals here. EgoScale measured a comparable law at roughly 20,000 hours, and recent work on zero-shot cross-embodiment transfer reaches overlapping conclusions from a different direction. Multiple groups converging on egocentric human video is a stronger market signal than any single company's report.
What reaching a million hours really does is open a larger design space. The next frontier is not 1M → 10M alone; it is that move on Axis I together with the Axis II move from basic egocentric observation toward state coverage, from hands toward complete physical interactions, and from human-specific recordings toward representations that survive a change of embodiment.
The first million hours settled whether scale works. The next ten million will settle what an hour is worth — and that is a question about collection design, as much as about scale.
Dyna-2 technical report: dyna.co/dyna-2
GENE-26.5 technical report: genesis.ai/blog/gene-26-5-advancing-robotic-manipulation-to-human-level