Keep on Going: Learning Robust Humanoid Motion Skills via Selective Adversarial Training