Autonomous drive tuning via Bayesian Optimization matches experts in minutes

Industrial electric drives require careful tuning of PID-like current control loops – a process that has traditionally demanded expert engineers and hours of manual adjustment. In a new paper submitted to IEEE ETFA 2026, David Petrovic and colleagues from the University of Padova introduce a fully automated commissioning approach using multi-objective Bayesian Optimization (BO). The drive is treated as a black-box system: no mathematical model or firmware modifications are needed. The optimizer iteratively updates controller parameters via closed-loop experiments on real hardware, directly measuring tracking error, time-weighted error, overshoot, and oscillatory behavior. By framing tuning as a multi-objective problem, the method reveals a Pareto front of viable trade-offs.

The team compared two BO strategies: Gaussian Process (GP)-based and Tree-structured Parzen Estimator (TPE)-based. While GP optimization can produce highly competitive final solutions, TPE proved better suited to the industrial setting. It converged faster, approximated the Pareto front more richly, and required less computational overhead – critical when evaluations are noisy, discrete parameters are involved, and budgets are limited. Experiments on a real motor drive (no-load conditions) showed that the autonomous system matched expert tuning performance within a few minutes, with no manual intervention. The approach also handled practical constraints like communication latency. This could dramatically reduce commissioning time and enable self-tuning drives in factories.