![[2602.13040] TCRL: Temporal-Coupled Adversarial Training for Robust Constrained Reinforcement Learning in Worst-Case Scenarios](https://arxiv.org/static/browse/0.3.4/images/arxiv-logo-fb.png){kind=logo}
[2602.13040] TCRL: Temporal-Coupled Adversarial Training for Robust Constrained Reinforcement Learning in Worst-Case Scenarios
Summary
The paper presents TCRL, a novel framework for robust constrained reinforcement learning that addresses challenges posed by temporally coupled adversarial perturbations, enhancing decision-making in safety-critical applications.
Why It Matters
As reinforcement learning is increasingly applied in safety-critical domains like autonomous driving and robotics, developing robust methods against adversarial attacks is crucial. TCRL's approach to modeling temporally coupled perturbations represents a significant advancement in ensuring safety and reliability in these applications.
Key Takeaways
- TCRL introduces a novel adversarial training framework for constrained reinforcement learning.
- It focuses on robustness against temporally coupled perturbations, which are often overlooked in existing methods.
- The framework employs a worst-case-perceived cost constraint to estimate safety costs effectively.
- Experimental results show TCRL outperforms traditional methods in various CRL tasks.
- The dual-constraint defense mechanism helps maintain reward unpredictability while countering adversarial attacks.
Computer Science > Machine Learning arXiv:2602.13040 (cs) [Submitted on 13 Feb 2026] Title:TCRL: Temporal-Coupled Adversarial Training for Robust Constrained Reinforcement Learning in Worst-Case Scenarios Authors:Wentao Xu, Zhongming Yao, Weihao Li, Zhenghang Song, Yumeng Song, Tianyi Li, Yushuai Li View a PDF of the paper titled TCRL: Temporal-Coupled Adversarial Training for Robust Constrained Reinforcement Learning in Worst-Case Scenarios, by Wentao Xu and Zhongming Yao and Weihao Li and Zhenghang Song and Yumeng Song and Tianyi Li and Yushuai Li View PDF HTML (experimental) Abstract:Constrained Reinforcement Learning (CRL) aims to optimize decision-making policies under constraint conditions, making it highly applicable to safety-critical domains such as autonomous driving, robotics, and power grid management. However, existing robust CRL approaches predominantly focus on single-step perturbations and temporally independent adversarial models, lacking explicit modeling of robustness against temporally coupled perturbations. To tackle these challenges, we propose TCRL, a novel temporal-coupled adversarial training framework for robust constrained reinforcement learning (TCRL) in worst-case scenarios. First, TCRL introduces a worst-case-perceived cost constraint function that estimates safety costs under temporally coupled perturbations without the need to explicitly model adversarial attackers. Second, TCRL establishes a dual-constraint defense mechanism on the reward t...
