Self-Improving Robust Preference Optimization