Significance extraction based on data augmentation for reinforcement learning

• SEQA improves generalization in visual reinforcement learning by masking interfering features and extracting salient features, enabling agents to focus on critical pixels. • The algorithm integrates data augmentation and a mask decoder network updated with critic losses, enhancing sample efficiency and stability during training. • SEQA outperforms state-of-the-art reinforcement learning methods in most DMControl-GB tasks, demonstrating superior generalization and sample efficiency. • The approach significantly boosts training efficiency and stability in DeepMind Control Suite environments, making it effective for real-world visual control applications.