Scattered data augmentation for generalization in visual reinforcement learning

doi:10.1016/j.neucom.2025.131492

Scattered data augmentation for generalization in visual reinforcement learning

2025-09-16

0

PRE

AI

DOI：10.1016/j.neucom.2025.131492

Original

Original request for help

Save

Abstract

En 中文

• We derive a generalization error upper bound for VRL from the perspective of distribution distance between training and test data. • We provide theoretical explanations for why DA can improve generalization in VRL and propose a novel DA framework ScDA, which treats the agent as the discriminator to ensure more divergent training data. • We conduct extensive experiments on DeepMind Control Generalization Benchmark2 and robotic tasks. The results demonstrate that ScDA can be effectively combined with baseline DA algorithms and significantly improve policy generalization, outperforming current SOTA approaches.

Scattered data augmentation for generalization in visual reinforcement learning

Abstract

Journal

Organization