“In zero-sum robotics self-play games, an adversarial opponent policy can be trained to reliably defeat a victim agent using seemingly random actions, even when trained for fewer than 3% of the time steps.”