The Economics of Robot Data: Cost per Demonstration Hour
Teleoperation, simulation, and egocentric video each carry a different cost per demonstration hour. Here is why that number drives robot-data strategy.
A person sits at a desk in a robotics lab, wearing a VR headset, folding the same towel for the two-hundredth time before lunch. Each fold takes a minute or two. On a good day that operator banks two or three hours of clean, labeled manipulation data across a full shift, and a fair share of what gets recorded is thrown away because the gripper slipped, the cloth bunched, or a camera clipped through the scene. That towel is quietly one of the most expensive materials in modern AI. It is priced not by the kilogram but by the demonstration hour.
Robot foundation models learn from demonstrations, and the scaling recipe the field borrowed from language models assumes data is cheap and close to infinite. Robot data is neither. Text is scraped; contact-rich manipulation is performed, one careful attempt at a time, on hardware that can cost as much as a car. So the number that actually governs strategy is not parameter count or context length. It is the marginal cost of the next usable demonstration hour, and nearly every serious lab is now organized, implicitly or explicitly, around driving that number down.
This piece walks through where the cost comes from, what the public datasets say about the scale still missing, and why the answer reshapes build-versus-buy decisions, simulation budgets, and even where a company chooses to incorporate.
Why a demonstration hour is the unit that matters
A demonstration hour is one hour of a robot, or a human stand-in, actually performing tasks while enough synchronized signal is recorded to train a policy: joint states, end-effector poses, contact forces, and time-aligned vision. It is a better accounting unit than trajectories or frames because it tracks the thing that is genuinely scarce, which is skilled human attention and physical setup time. You can inflate a frame count by lowering the shutter speed. You cannot fake an hour of an expert solving contact-rich tasks.
The asymmetry with language is stark. A frontier text model trains on trillions of tokens harvested at almost no marginal cost. The largest open robot-learning effort to date, Open X-Embodiment, pooled more than a million real trajectories across 22 robot embodiments from over 20 institutions, and that global aggregate is still tiny next to the text corpora behind any modern LLM. When supply is that thin, the price of each additional hour sets the pace of the whole field.
What the public datasets reveal about scale
The open datasets are useful precisely because they publish their own logistics. DROID gathered roughly 76,000 teleoperated trajectories, on the order of 350 hours of interaction, across 564 scenes and 13 institutions over about 12 months. RH20T assembled more than 110,000 contact-rich manipulation sequences with paired vision, force, and audio. Both are serious, multi-year, multi-lab undertakings, and both still measure their interaction time in hundreds of hours, not millions.
That is the quiet lesson. Even a well-funded consortium, running many rigs in parallel, tops out at a throughput that looks nothing like web scraping. The ceiling is set by human operators and by the physical world running in real time. You cannot parallelize a towel fold past the number of hands and arms you actually own.
The cost stack: capital, labor, throughput, yield
The cost per usable demonstration hour is a short equation with four terms. Capital is the amortized cost of the robot, sensors, and rig per hour of use. Labor is the operator wage, plus the annotators and quality reviewers behind them. Throughput is how many good demonstrations one station produces per hour. Yield is the fraction that survives review; contact-rich tasks fail often, and failed or corrupted takes are dropped. Put simply, cost per usable hour equals capital-per-hour plus labor-per-hour, all divided by yield.
The yield term is the one newcomers underestimate. If a third of your takes are unusable, your effective cost rises by half before anyone has trained a model. Teleoperation labor alone lands, very roughly, on the order of tens of dollars per operator-hour in a mid-cost region, and that is before capital amortization and before yield losses. The table below sketches the order-of-magnitude tradeoffs between the common sources. These are directional, not vendor quotes.
| Source | Relative cost per usable hour | Throughput | Embodiment match | Main constraint |
|---|---|---|---|---|
| Teleoperated real robot | Highest | Low | Exact | Labor-bound, real-time |
| Pooled multi-lab robot data | Amortized, lower per org | Medium | Mixed | Heterogeneous hardware |
| Simulation | Lowest marginal | Very high | Approximate | Sim-to-real gap |
| Egocentric human video | Low to moderate | High | Human, needs retargeting | No native action labels |
The bottleneck in robot learning is no longer the model. It is the marginal cost of the next honest hour of contact-rich demonstration.
Egocentric video changes the arithmetic
The most interesting move on the cost curve is to stop paying for the robot during collection. If a human simply performs the task while wearing head-mounted cameras and body sensors, you decouple data capture from both expensive hardware and slow teleoperation. Ego-Exo4D captured roughly 1,286 hours of paired first-person and third-person video of skilled activities, and the earlier Ego4D effort released on the order of 3,670 hours of egocentric footage. Recording a human doing a task is far cheaper per hour than steering a robot through it.
The catch is the embodiment gap. Human video has no joint torques, no gripper states, and a body that does not match the robot, so the frames need retargeting into a robot-usable action space before a policy can consume them. Cross-embodiment transfer is exactly the problem groups like Physical Intelligence are chipping at. Egocentric capture does not erase the cost of a good demonstration; it moves the cost from real-time teleoperation into upstream sensing and downstream retargeting, where more of it can be engineered and reused.
How cost per hour drives strategy
Once you take the number seriously, the strategic choices fall out of it. If your marginal cost is tens of dollars per usable hour and a capable policy needs thousands of hours per skill family, data is a capital line, not a rounding error, and it belongs in the same conversation as compute. Pooling, in the Open X-Embodiment style, cuts the per-organization cost but trades away exclusivity and clean embodiment match. Simulation converts operating cost into a sim-to-real risk you pay down later. Egocentric capture lowers the per-hour price but raises the engineering bill for retargeting.
Then there is the part that lives in the policy cluster: who owns the data, who consented, and under which jurisdiction it was captured. The EU Data Act reshapes access and sharing rights over machine-generated data, and the EU AI Act attaches documentation and governance duties to the training data behind high-risk systems. Both raise the fixed cost of doing data collection properly, and both make clean provenance a durable asset rather than a compliance afterthought. A demonstration hour with clear consent and a defensible jurisdiction is worth more than a cheaper hour with a murky origin, because only one of them is safe to ship inside a model a partner will deploy.
None of this shows up in a benchmark chart, which is why it is easy to ignore until the invoices land. The teams that will matter over the next few years are not only the ones with the cleverest architectures. They are the ones that turn the demonstration hour into something they can produce reliably, document honestly, and drive down the cost of, quarter after quarter. The towel folds one attempt at a time, and someone still has to pay for every one of them.