“A properly designed robotic foundation model should not hallucinate because each prediction is grounded in a new, direct observation of the physical world state, unlike language models where errors can compound.”