Abstract

Controller parameter tuning in automotive development typically requires extensive real-world testing by application engineers, which is time-consuming, costly, and risks evaluating parameter configurations that lead to poor or unsafe closed-loop behavior. In industrial practice, controller design follows a two-stage workflow – initial design and tuning in simulation (low-fidelity), followed by application and refinement on the real vehicle (high-fidelity) – placing additional requirements on tuning methods. Two key challenges arise in this setting: reducing the number of costly real-world experiments, and avoiding unsafe parameter configurations. In this work, we propose a framework for cautious learning of controller parameters based on constrained multi-fidelity Bayesian optimization, explicitly combining multi-fidelity modeling with safety-related constraints. The proposed framework guides the optimization toward promising regions while avoiding parameter choices that may degrade performance or violate operational limits. This enables efficient and risk-averse exploration even when high-fidelity evaluations are scarce and expensive. Simulation results show that the framework achieves an effective trade-off between learning efficiency and avoidance of undesirable configurations, offering a practical pathway toward automated, data-efficient controller tuning in industrial applications.

Link: https://ieeexplore.ieee.org/document/11625264