Safe learning-based control via Predictive Safety Filters

With the increasing complexity of modern systems, it remains challenging for controllers to handle complicated, non-convex cost functions while still guaranteeing safety and stability. Classical control methods—such as Model Predictive Control (MPC)—provide rigorous analyses for stability and constraint satisfaction, yet they often limit the joint design of rich cost functions and stability properties. While recent learning-based control approaches can optimise performance, their black-box nature makes satisfying hard constraints difficult. In this work, we develop and validate a learning-based control algorithm that embeds a Predictive Safety Filter (PSF) to deliver formal guarantees of stability and constraint satisfaction while retaining suboptimal performance. We demonstrate its effectiveness on a single-pendulum stabilization task with obstacle avoidance, showing that safety and stability can be maintained without sacrificing practical performance.

Join at: imt.lu/aula2