Imagine to Ensure Safety in Hierarchical Reinforcement Learning
Keywords: safe reinforcement learning, machine learning, model based reinforcement learning, hierarchical reinforcement learning
Abstract: This work investigates the safe exploration problem, where an agent must maximize performance while satisfying safety constraints. To address this problem, we propose a method that includes a learnable world model and two policies, a high-level policy and a low-level policy, that ensure safety at both levels. The high-level policy generates safe subgoals for the low-level policy, which progressively guide the agent towards the final goal. Through trajectory imagination, the low-level policy learns to safely reach these subgoals. The proposed method was evaluated on the standard benchmark, SafetyGym, and demonstrated superior performance quality while maintaining comparable safety violations compared to state-of-the-art approaches. In addition, we investigated an alternative implementation of safety in hierarchical reinforcement learning (HRL) algorithms using Lagrange multipliers, and demonstrated in the custom long-horizon environments SafeAntMaze that our approach achieves comparable performance while more effectively satisfying safety constraints, while the flat safe policy fails to accomplish this task.
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: 14037
Loading