First-person video and depth captured from the worker's own point of view, so the model sees the task exactly as the person performing it does.

Body and hand pose tracked over time, turning a movement into a measurable trajectory a policy can imitate.

Live readings from the equipment itself: controller signals, loads, speeds and alarms, time-synced to what the human is doing.

What the task was, where it happened, which objects were involved and what changed around it. Context that makes an action interpretable.

Whether the attempt worked, how it failed, and how it was recovered. Failure and correction are labeled, not discarded.

Consent, usage rights and a traceable chain of custody shipped with every episode, so legal clearance takes days instead of quarters.