“OpenAI trained a single goal-conditioned policy that can zero-shot generalize to unseen robotic manipulation tasks using an asymmetric self-play framework.”