ASOR: Anchor State Oriented Regularization for Policy Optimization under Dynamics Shift

Zhenghai Xue, Lang Feng, Jiacheng Xu, Kang Kang, Xiang Wen, Bo An, Shuicheng YAN

Keywords: Reinforcement Learning, Policy Transfer, Dynamics Shift, Policy Regularization

Abstract:

To train neural policies in environments with diverse dynamics, Imitation from
Observation (IfO) approaches aim at recovering expert state trajectories. Their
success is built upon the assumption that the stationary state distributions induced
by optimal policies remain similar despite dynamics shift. However, such an
assumption does not hold in many real world scenarios, especially when certain
states become inaccessible during environment dynamics change. In this paper,
we propose the concept of anchor states which appear in all optimal trajectories
under dynamics shift, thereby maintaining consistent state accessibility. Instead of
direct imitation, we incorporate anchor state distributions into policy regularization
to mitigate the issue of inaccessible states, leading to the ASOR algorithm. By
formally characterizing the difference of state accessibility under dynamics shift,
we show that the anchor state-based regularization approach provides strong lower-
bound performance guarantees for efficient policy optimization. We perform
extensive experiments across various online and offline RL benchmarks, including
Gridworld, MuJoCo, MetaDrive, D4RL, and a fall-guys like game environment,
featuring multiple sources of dynamics shift. Experimental results indicate ASOR
can be effectively integrated with several state-of-the-art cross-domain policy
transfer algorithms, substantially enhancing their performance.

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: 2284