“OpenAI used an asymmetric self-play framework with two policies, "Alice" for goal generation and "Bob" for goal solving, to train a robot for tabletop rearrangement tasks.”