Combined data augmentation framework for generalizing deep reinforcement learning from pixels

Xi Xiong, Chun Shen, Junhong Wu, Shuai Lü, Xiaodan Zhang

Abstract: Highlights•We reconsider the characteristics of applying data augmentation in visual RL.•We design a framework to use all data augmentation operators simultaneously.•We enhance diversity and reduce hardness through variant combinations.•Our method outperforms SOTA on most DMControl tasks.