How Much Data a Humanoid Skill Needs: Sample Efficiency

A diffusion policy can learn to pull a bagel out of a toaster from roughly one hundred demonstrations. Give it a fresh kitchen, a different toaster, and afternoon light instead of morning, and the same model, trained on the same hundred demonstrations, will miss. Nothing about the skill changed. What changed is the slice of the world you asked the skill to survive.

Ask ten robotics engineers how much data a humanoid skill needs and you will get answers spanning three orders of magnitude. Fifty teleoperated episodes. A few hours. Ten thousand hours. All of them are defensible, and that is the problem. "Demonstrations per skill" is a number with a hidden denominator, and until you fix the denominator, the number means almost nothing.

This piece tries to give an honest range instead of a single figure, and to name what moves you along it: task complexity, how much environmental variation you want to cover, and whether you start from a generalist model or from scratch. The short version is that a skill in a fixed setup is cheap, and the same skill that holds up in the open world is not.

The unit problem

A demonstration is a trajectory: a sequence of observations and actions that completes a task once. But "pick up the cup" collapses an enormous range of difficulty into four words. Pick up one cup, in one pose, on one table, under one light, and a modern imitation-learning stack needs surprisingly little. Pick up any cup, anywhere, and you are no longer learning a skill; you are learning a distribution.

So the useful question is not how many demonstrations a skill needs, but how many it needs to reach a target reliability across a target range of conditions. Contact-rich tasks (peeling, plugging, folding) cost more than free-space reaches. Tight-tolerance tasks cost more than forgiving ones. And every axis of variation you add, new objects, new backgrounds, new lighting, new starting states, multiplies the demonstrations required to cover it rather than adding to them.

The question is never how many demonstrations a skill needs. It is how much of the world you want the skill to survive.

What single-task imitation actually costs

Start with the cheapest regime: one task, one setup, trained from scratch. Here the literature is remarkably consistent. Action Chunking Transformers on the ALOHA platform learn contact-rich bimanual tasks from on the order of fifty demonstrations. Diffusion Policy reports reliable single-task performance in roughly the 100-200 demonstration range, sometimes fewer for simple motions. Google's RT-1 collected about 130,000 episodes across more than 700 tasks, which averages to a couple hundred demonstrations per task even at fleet scale.

So a first honest anchor: for a well-scoped manipulation skill in fixed conditions, budget tens to low hundreds of demonstrations, a few hours of teleoperation. That is also the number most often quoted out of context, because it describes the easiest version of the problem and quietly assumes the test looks like the training set.

The pretraining discount

Those numbers assume you train each skill in isolation. Almost nobody serious does that anymore. The dominant recipe is to pretrain a generalist policy on a large, diverse corpus, then fine-tune on a modest number of demonstrations for the target skill. Pretraining pays down the cost of every new skill afterward.

The corpora are now substantial. Open X-Embodiment aggregates more than a million real trajectories across 22 robot embodiments and hundreds of skills. DROID adds roughly 76,000 trajectories collected in the wild across 564 scenes. Physical Intelligence's pi-zero flow model and NVIDIA's GR00T N1 are both vision-language-action models pretrained on this kind of cross-embodiment mixture, and both are designed to be adapted to a new task with far less data than training it cold would take. Toyota Research Institute's Large Behavior Model work makes the mechanism explicit: co-training a single network on hundreds of tasks improves per-task performance and cuts the demonstrations needed to add the next one.

The discount is real but not free. Fine-tuning still needs demonstrations in the target embodiment and environment, and the further your robot and setting sit from the pretraining distribution, the more you need. A humanoid hand is not a parallel-jaw gripper; much of the cheapest existing robot data does not transfer cleanly to it.

Why human video changes the arithmetic

Teleoperated robot data has a hard ceiling: every demonstration needs a robot and an operator, so the corpus grows one arm-hour at a time. Human demonstration does not. Ego4D assembled roughly 3,670 hours of first-person daily activity; Ego-Exo4D added more than 1,200 hours of skilled tasks captured from synchronized first- and third-person views. That is video of hands doing real work at a scale no teleoperation fleet will match soon.

The catch is the embodiment gap. Human video carries the priors that matter, how a task decomposes, where contact happens, how objects behave, but no robot joint commands and often no reliable 3D. Converting it into supervision a policy can train on, retargeted hands, recovered contact, action labels, is the active research frontier, and it is where a lot of the field's leverage now sits. GR00T's training mixture already blends real robot data, simulation, and human video precisely because none of the three is sufficient alone.

Budgeting data for a new skill

Put the pieces together and a practical rule emerges. To estimate the data for a new humanoid skill, do not ask for a single number; specify three things and read off a range.

• Reliability target. A 70% demo-day success rate and a 99% production rate are separated by a long tail that costs most of the data.
• Variation target. One fixed cell, or any kitchen? Each independent axis of variation roughly multiplies the requirement.
• Starting point. From scratch, budget tens to hundreds of demonstrations for the fixed case. From a strong pretrained generalist, often less for that same fixed case, but the open-world case still runs into the thousands.

As a rough shape: fixed-setup skills live in the tens to hundreds of demonstrations; the same skill hardened for real deployment across homes or stores tends to climb into the thousands of demonstrations, or the equivalent hundreds of hours, most of it spent not on the motion but on the long tail of conditions around it.

The honest number

So, how much data does a humanoid skill really need? Tens of demonstrations to look good once, hundreds to work in a fixed cell, and thousands, or the human-hour equivalent, to survive contact with a world it has not seen. The figure is a function, not a constant, and its steepest input is generalization. Anyone who quotes you one number for data per skill is really quoting you their denominator, and hoping you will not ask.