Forget YouTube videos—frontier physical AI models need multiple camera angles, dense annotation, and soon, brain wave readings.