A video can show that work happened without showing an AI system how the work was done. Physical training data needs the action, the state around it, the reason for each decision, and the result that followed.
This note describes the collection principles we use to turn real work into structured learning material. It presents a framework, not a model benchmark.
A task is more than a clip
Useful capture keeps the demonstration aligned with synchronized context. Camera views, tool state, timing, instructions, and outcomes need a shared clock. Without that alignment, important details become guesswork during training.
Variation needs a reason
More footage is not automatically more coverage. Variation should map to changes that affect the task, including operator technique, environment, object state, tool choice, and failure mode. Each dimension needs a reason to exist in the collection plan.
Review belongs inside collection
Quality review works best when it can send precise feedback back to capture teams. Missing state, unclear intent, unsafe execution, and incomplete outcomes can then be corrected before they become permanent gaps in a dataset.
Figures use illustrative values.