Bayesian Optimization for Auto-tuning Cascaded PID Control of an (e)ROV
From manually tuned PID control to a Bayesian Optimized auto-tuned variant
D.H.J. Poutsma (TU Delft - Mechanical Engineering)
L. Ferranti – Mentor (TU Delft - Mechanical Engineering)
Vasso Reppa – Graduation committee member (TU Delft - Mechanical Engineering)
N.Y. Sheth – Graduation committee member (Fugro)
More Info
expand_more
Other than for strictly personal use, it is not permitted to download, forward or distribute the text or part of it, without the consent of the author(s) and/or copyright holder(s), unless the work is under an open content license such as Creative Commons.
Abstract
Industrial electric Remotely Operated Vehicles (eROVs) rely on cascaded PI controllers whose gains must be re-tuned manually whenever payload, manipulator, or thruster configuration changes—a recurring and costly operational bottleneck. Existing auto-tuning approaches for cascaded controllers have been demonstrated only on linear, well-modelled, single-axis systems, leaving the combination of nonlinear multi-DOF dynamics, constrained thruster allocation, direction-dependent asymmetry, and tether-induced disturbances unaddressed.
This thesis presents a data-driven auto-tuning workflow for cascaded PI control on industrial eROVs. The workflow incorporates a standardised evaluation protocol, safety screening mechanisms, a composite cost function capturing multiple closed-loop performance criteria, and a hyperparameter configuration that generalises across axes and platforms without per-axis adjustment. It is validated on two physically distinct platforms differing by a factor of approximately 4.6 in mass: one in high-fidelity simulation (four independent tuning runs) and one on real hardware at an operational test site (two runs). Across all runs, the auto-tuned gains consistently match or outperform the manually tuned baseline on every axis where the evaluation protocol provides informative signals, with substantial cost reductions observed on both platforms. Residual cases where the workflow does not improve upon the baseline are traced to identifiable limitations in the evaluation protocol rather than in the optimiser itself. The auto-tuned gains for several axes have been transferred to the production vehicle, and independent pilot assessments confirm clearly improved closed-loop behaviour.
The workflow is ready for industrial deployment as a drop-in replacement for manual cascaded PI tuning on eROVs.
help_outline