Which properties make augmentation useful for visual RL? This study motivates Random PadResize, which introduces spatial variation with limited distortion, and Cycling Augmentation, which alternates augmentation types over time. Experiments cover DeepMind Control and CARLA. Read the conference record.
NeurIPS · Neural Information Processing Systems · 2023
Learning Better with Less: Effective Augmentation for Sample-Efficient Visual Reinforcement Learning
Guozheng Ma, Linrui Zhang, Haoyu Wang, Lu Li, Zilin Wang, Zhen Wang, Li Shen, Xueqian Wang, Dacheng Tao
Related research
Learning from a different view
Useful visual diversity depends on both the transformation and the way it enters training.