[Title] Fourier Guided Adaptive Adversarial Augmentation for Generalization in Visual Reinforcement Learning
[Journal] Annual AAAI Conference on Artificial Intelligence (AAAI) 2025.
[Authors] Jeongwoon Lee, and Hyoseok Hwang*
[Summary] We introduces Fourier Guided Adaptive Adversarial Augmentation (FGA3), a new augmentation technique designed to improve the robustness of Visual Reinforcement Learning (RL) agents when faced with unseen environments. FGA3 addresses the domain gap between training and testing by focusing on two key aspects: Frequency Domain Style Augmentation & Adaptive Adversarial Perturbation
[Key Figure]
Framework of FGA3
[Key Result]Comparison with other methods on Random Colors and Natural Videos benchmark in DMControl-GB after training on 500K frames. We provide the mean and standard deviation of episode return trained with three different random seeds. (·) represents the standard deviation.