Teams building vision-language-action (VLA) models tend to measure data in hours. It is a convenient unit, but it hides an uncomfortable fact: two hours of recorded human activity can differ by an order of magnitude in how much they actually move a model. An hour of footage is raw material. A usable demonstration hour is a finished input: segmented, labeled, signal-rich and cleared for use. The distance between the two is where most of the work lives.
Here is how we think about that distance, and what has to be true before an egocentric demonstration hour is worth putting in front of a policy.
Real-world diversity, not repetition
The first failure mode is sameness. An hour recorded in one kitchen, with one pair of hands, under one lighting condition, teaches a model to succeed in exactly that setting and nowhere else. Usable hours carry variation the model will meet in deployment: different environments, object instances, grips, lighting, clutter and recovery from small mistakes. Diversity is not a nice-to-have that you sprinkle on afterward; it is what lets a policy generalize instead of memorize. We commission collection against a task spec precisely so that the variation is deliberate rather than accidental.
Clean segmentation
Raw footage is continuous; learning is not. A demonstration only becomes tractable once it is cut into coherent episodes: a single attempt at a single task, with a clear start and end. Clean segmentation means the boundaries fall in the right place, idle and setup time is trimmed, and a "one hour" figure reflects demonstration content rather than the collector walking between rooms. Poorly segmented data quietly poisons a batch: the model spends capacity modeling dead time and transitions that were never meant to be learned.
Pose and contact signals
Pixels alone under-specify what a hand is doing. For manipulation, the moments that matter are contact events: when a grasp closes, when force is applied, when an object is released. A usable hour carries the signals that make those moments legible: hand and finger pose, object pose where it can be recovered, and contact or grasp state aligned to the video timeline. These channels let what let a model connect an intention to the physical change it produced, rather than guessing from appearance.
Action labels that match the frames
An episode needs to say what was being attempted and whether it succeeded. Action labels (task descriptions, sub-step boundaries, success and failure markers) turn footage into supervision. The requirement is not just that labels exist but that they are time-aligned and consistent: the same event described the same way across collectors and sessions. Inconsistent labeling caps the ceiling of what any downstream model can learn, no matter how many hours sit behind it.
Rights clearance
A demonstration hour is only usable if you are actually allowed to train on it. That means documented consent from the person who recorded it, clearance for the setting and any incidental people or property in frame, and a provenance trail that ties each clip to its consent and collection terms. Rights are not a legal footnote to attach at delivery; they are a property of the data itself. An hour without a clean chain is a liability regardless of how good the footage looks, and we treat missing provenance as disqualifying rather than fixable after the fact.
Quality control as the gate
Every property above can degrade, and the only way to keep an aggregate corpus trustworthy is to check each hour before it counts. Our quality review combines automated checks (coverage of the task spec, signal completeness, label alignment, provenance presence) with human review of the harder judgments a script cannot make reliably. Footage that captures identifiable bystanders, children or private spaces against our collection guidelines is rejected at this stage rather than patched later. QC is where "recorded" becomes "usable," and it is the reason a delivered hour behaves like the spec it was ordered against.
The refinement pipeline
None of these properties appear on their own. Diversity comes from how collection is commissioned; segmentation, pose and contact signals, and action labels come from refinement; rights and QC come from the review and provenance layer that runs through the whole process. That end-to-end path, from a task spec, through the collector community, through refinement and review, to a rights-cleared delivery in your format, is what a usable demonstration hour actually is. The hour count on an invoice is just the last visible number; the work that makes it mean something happens before it ever reaches you.
If you are evaluating demonstration data for a VLA program, the useful question is never how many hours a vendor can produce. It is how many of those hours would survive the checks above. That is the number worth optimizing.