Blog

[Summary]
Reinforcement learning (RL) has proven its potential in complex decision-making tasks. Yet, many RL systems rely on manually crafted state representations, requiring effort in feature engineering. Visual Reinforcement Learning (VRL) offers a way to address this challenge by enabling agents to learn directly from raw visual input. Nonetheless, VRL continues to face generalization issues, as models often overfit to specific domain features. To tackle this issue, we propose Diffusion Guided Adaptive Augmentation (DGA2), an augmentation method that utilizes Stable Diffusion to enhance domain diversity. We introduce an Adaptive Domain Shift strategy that dynamically adjusts the degree of domain shift according to the agent’s learning progress for effective augmentation with Stable Diffusion. Additionally, we employ saliency as the mask to preserve the semantics of data. Our experiments on the DMControl-GB, Adroit, and Procgen environments demonstrate that DGA2 improves generalization performance compared to existing data augmentation and generalization methods.Deep learning has advanced vision tasks such as classification, segmentation, and detection. However, in real-world scenarios, models often encounter domains that differ from the ones seen during training, which can lead to substantial performance degradation. To mitigate the effects of distribution shifts, domain generalization (DG) aims to enable models to generalize effectively to unseen target domains. Recent DG approaches use generative models like diffusion models to augment data with text prompts. However, these methods rely on domain-specific textual inputs and require costly fine-tuning, which limits their scalability. We propose TRIDENT, a framework that overcomes these limitations by eliminating the need for text prompts and leveraging the linear structure of CLIP embeddings. TRIDENT decomposes image embeddings into three components—domain, class, and attribute—enabling precise control over semantic content. By reassembling each embedding component, we generate semantically valid and structurally coherent synthetic samples across domains. This allows efficient and diverse data synthesis without retraining diffusion models. TRIDENT operates through lightweight embedding-space manipulation, significantly reducing computational overhead. Extensive experiments on standard DG benchmarks (e.g., PACS, VLCS, and OfficeHome) demonstrate that TRIDENT achieves competitive or superior performance compared with existing approaches. Furthermore, qualitative evaluations and comprehensive analyses confirm that TRIDENT not only enables efficient and diverse data synthesis but also demonstrates the effectiveness of the proposed decomposition strategy.