From learning to safety: A Direct Data-Driven Framework for Constrained Control
Ensuring safety in the sense of constraint satisfaction for learning-based control is a critical challenge, especially in the model-free case. While safety filters address this challenge in the model-based setting by modifying unsafe control inputs, they typically rely on predictive models derived from physics or data. This reliance limits their applicability for advanced model-free learning control methods. To address this gap, we propose a new optimization-based control framework that determines safe control inputs directly from data. The benefit of the framework is that it can be updated through arbitrary model-free learning algorithms to pursue optimal performance. As a key component, the concept of direct data-driven safety filters (3DSF) is first proposed. The framework employs a novel safety certificate, called the state-action control barrier function (SACBF). We present three different schemes to learn the SACBF. Furthermore, based on input-to-state safety analysis, we present the error-to-state safety analysis framework, which provides formal guarantees on safety and recursive feasibility even in the presence of learning inaccuracies. The proposed control framework bridges the gap between model-free learning-based control and constrained control, by decoupling performance optimization from safety enforcement. Simulations on vehicle control illustrate the superior performance regarding constraint satisfaction and task achievement compared to model-based methods and reward shaping.