COSTAR: Dynamic Safety Constraints Adaptation in Safe Reinforcement Learning

Xiuhui Zhang, Xingyu Jiang, Ning Gao, Mukang You, Yue Deng

Keywords: Reinforcement Learning

TL;DR: We introduce the COSTAR framework for safe reinforcement learning, enhancing the ability of agents to adapt to dynamic safety constraints, including variable cost functions and safety thresholds.

Abstract: Recent advancements in safe reinforcement learning (safe RL) have focused on developing agents that maximize rewards while satisfying predefined safety constraints. However, the challenge of learning policies capable of generalizing to dynamic safety requirements has rarely been explored. To this end, we propose a novel COntrastive Safe TAsk Representation (COSTAR) framework for safe RL, which can boost existing algorithm's generalization to dynamic safety constraints, including variable cost functions and safety thresholds.In COSTAR, we employ a Safe Task Encoder to extract safety-specific representations from trajectory contexts, effectively distinguishing between various safety constraints with contrastive learning. It is noteworthy that our framework can integrate with existing safe RL algorithms and possesses zero-shot adaptation capability to varying safety constraints during deployment. Extensive experiments demonstrate that our COSTAR framework consistently achieves high rewards while maintaining low costs, and exhibits robust generalization capabilities when dealing with out-of-distribution (OOD) tasks.

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.

Reciprocal Reviewing: I understand the reciprocal reviewing requirement as described on https://iclr.cc/Conferences/2025/CallForPapers. If none of the authors are registered as a reviewer, it may result in a desk rejection at the discretion of the program chairs. To request an exception, please complete this form at https://forms.gle/Huojr6VjkFxiQsUp6.

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: 9481