RDT 2, the sequel to RDT-1B, is the first foundation model that achieves zero-shot deployment on unseen embodiments for simple open-vocabulary tasks.