Learning over Forward-Invariant Policy Classes: Reinforcement Learning without Safety Concerns
Fuente:
arXiv
Gespeichert in:
| Hauptverfasser: | , , , |
|---|---|
| Format: | Preprint |
| Veröffentlicht: |
2026
|
| Schlagworte: | |
| Online-Zugang: | |
| Tags: |
Tag hinzufügen
Keine Tags, Fügen Sie den ersten Tag hinzu!
|
| _version_ | 1866911578265223168 |
|---|---|
| author | Tsai, Chieh Zahed, Muhammad Junayed Hasan Hariri, Salim Rastgoftar, Hossein |
| author_facet | Tsai, Chieh Zahed, Muhammad Junayed Hasan Hariri, Salim Rastgoftar, Hossein |
| contents | This paper proposes a safe reinforcement learning (RL) framework based on forward-invariance-induced action-space design. The control problem is cast as a Markov decision process, but instead of relying on runtime shielding or penalty-based constraints, safety is embedded directly into the action representation. Specifically, we construct a finite admissible action set in which each discrete action corresponds to a stabilizing feedback law that preserves forward invariance of a prescribed safe state set. Consequently, the RL agent optimizes policies over a safe-by-construction policy class. We validate the framework on a quadcopter hover-regulation problem under disturbance. Simulation results show that the learned policy improves closed-loop performance and switching efficiency, while all evaluated policies remain safety-preserving. The proposed formulation decouples safety assurance from performance optimization and provides a promising foundation for safe learning in nonlinear systems. |
| format | Preprint |
| id |
arxiv_https___arxiv_org_abs_2604_07875 |
| institution | arXiv |
| publishDate | 2026 |
| record_format | arxiv |
| spellingShingle | Learning over Forward-Invariant Policy Classes: Reinforcement Learning without Safety Concerns Tsai, Chieh Zahed, Muhammad Junayed Hasan Hariri, Salim Rastgoftar, Hossein Systems and Control This paper proposes a safe reinforcement learning (RL) framework based on forward-invariance-induced action-space design. The control problem is cast as a Markov decision process, but instead of relying on runtime shielding or penalty-based constraints, safety is embedded directly into the action representation. Specifically, we construct a finite admissible action set in which each discrete action corresponds to a stabilizing feedback law that preserves forward invariance of a prescribed safe state set. Consequently, the RL agent optimizes policies over a safe-by-construction policy class. We validate the framework on a quadcopter hover-regulation problem under disturbance. Simulation results show that the learned policy improves closed-loop performance and switching efficiency, while all evaluated policies remain safety-preserving. The proposed formulation decouples safety assurance from performance optimization and provides a promising foundation for safe learning in nonlinear systems. |
| title | Learning over Forward-Invariant Policy Classes: Reinforcement Learning without Safety Concerns |
| topic | Systems and Control |
| url | https://arxiv.org/abs/2604.07875 |