[2503.08751] Disentangled World Models: Learning to Transfer Semantic Knowledge from Distracting Videos for Reinforcement Learning

Computer Science > Computer Vision and Pattern Recognition arXiv:2503.08751 (cs) [Submitted on 11 Mar 2025 (v1), last revised 4 Apr 2026 (this version, v3)] Title:Disentangled World Models: Learning to Transfer Semantic Knowledge from Distracting Videos for Reinforcement Learning Authors:Qi Wang, Zhipeng Zhang, Baao Xie, Xin Jin, Yunbo Wang, Shiyu Wang, Liaomo Zheng, Xiaokang Yang, Wenjun Zeng View a PDF of the paper titled Disentangled World Models: Learning to Transfer Semantic Knowledge from Distracting Videos for Reinforcement Learning, by Qi Wang and 8 other authors View PDF HTML (experimental) Abstract:Training visual reinforcement learning (RL) in practical scenarios presents a significant challenge, $\textit{i.e.,}$ RL agents suffer from low sample efficiency in environments with variations. While various approaches have attempted to alleviate this issue by disentangled representation learning, these methods usually start learning from scratch without prior knowledge of the world. This paper, in contrast, tries to learn and understand underlying semantic variations from distracting videos via offline-to-online latent distillation and flexible disentanglement constraints. To enable effective cross-domain semantic knowledge transfer, we introduce an interpretable model-based RL framework, dubbed Disentangled World Models (DisWM). Specifically, we pretrain the action-free video prediction model offline with disentanglement regularization to extract semantic knowledge ...