That difference—not model size or novelty—is what makes embodied AI harder to deploy and more consequential when it fails.