A humanoid is not a robot arm with legs attached. The five kinds of data a humanoid policy needs that arms, quadrupeds, and mobile manipulators never do, and which of them simulation can actually carry.
A one-time cleanup will not keep a robot dataset good. This is a six-stage loop for raising the quality of a dataset you already have: audit, score, version, remediate, re-validate, and monitor, with the metrics to track at each stage.
Combining sensors, video, motion, and language into one physical AI dataset is not a capture problem, it is an alignment problem. Here is the five-stage breakdown of where multimodal programs actually fail, with real numbers from the research.
A robot demonstration is not a video clip. It is a structured bundle of synchronized observations, action labels, phase boundaries, and outcome flags built to be consumed by a training pipeline, not watched. Here is what is actually inside one.
Physical AI vs embodied AI, explained: where the terms overlap, where they split, and why the difference changes your robot data plan and vendor choices.