Don't Touch What Matters: Task-Aware Lipschitz Data Augmentation for Visual Reinforcement Learning