AsymDreamer: Safe Reinforcement Learning From Pixels with Privileged World Models

Dongchi Huang, Kaige Zhang, Yang Li, Yi Zhan, Chunhe Xia

Keywords: Safe Reinforcement Learing; World Model

Abstract:

Safe Reinforcement Learning from partial observations frequently struggles with rapid performance degradation and often fails to satisfy safety constraints. Upon deeper analysis, we attribute this problem to the lack of necessary information in partial observations and inadequate sample efficiency. World Models can help mitigate this issue, as they offer high sample efficiency and the capacity to memorize historical information. In this work, we introduce AsymDreamer, an approach based on the Dreamer framework that specializes in exploiting low-dimensional privileged information to build world models, thereby enhancing the prediction capability of critics. To ensure safety, we employ the Lagrangian method to incorporate safety constraints. Additionally, we formulate our approach as an Asymmetric CPOMDPs (ACPOMDPs) framework and analyze its superiority compared to the standard CPOMDP framework. Various experiments conducted on the Safety-Gymnasium benchmark demonstrate that our approach outperforms existing approaches dramatically in terms of performance and safety.

Primary Area: reinforcement learning

Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.

Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide.

Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors’ identity.

No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.

Submission Number: 9176