Model-Free Learning Reference Governor With Enhanced Data Collection for Safety-Critical Control Systems

Ensuring constraint satisfaction in control systems without relying on accurate models is essential for real-world applications. Learning-based Reference Governors (LRGs) address this challenge by leveraging data-driven adaptation to improve constraint handling. However, existing safety-critical LRG methods often suffer from slow learning speeds and require full state measurements. This paper presents an enhanced safety-critical LRG framework that accelerates learning by redefining the peak deviation function and proposes an output-only measurement version that increases its practical applicability. A case study on a spacecraft with a flexible appendage demonstrates the effectiveness of the approach, showing improved learning speed and closed-loop convergence.