A control-barrier-function-based algorithm for policy adaptation in reinforcement learning

• Paper
• Code and data

Abstract

This paper studies adaptation of a predesigned policy from an original objective to a trade-off objective while explicitly constraining deviation from original performance. The method builds a closed-loop dynamics for policy parameters and uses control barrier functions to enforce constraint satisfaction during adaptation. Experiments on Cartpole, Lunar Lander, and a quadruped robot show practical and safe policy adaptation.

Demo

Citation

Hao, Wenjian, Zehui Lu, Nicolas Miguel, and Shaoshuai Mou. 2025. "A Control-Barrier-Function-Based Algorithm for Policy Adaptation in Reinforcement Learning." arXiv preprint arXiv:2510.02720.

@techreport{WHao2025cbfpolic,
  author = {W Hao, Z Lu, N Miguel, S Mou},
  year = {2025},
  title = {A Control-Barrier-Function-Based Algorithm for Policy Adaptation in Reinforcement Learning},
  number = {arXiv:2510.02720},
  url = {https://arxiv.org/pdf/2510.02720}
}