“The policy trained with asymmetric self-play struggled with challenging tasks such as goalball capture, the rainbow task, and stacking more than 3 blocks.”