Data-enabled Predictive Control of Robotic Systems

Abstract

Autonomous driving and highly automated manufacturing processes are only two examples for rapidly increasing system complexity. At the same time, almost unlimited connectivity produces never before seen amounts of data. As a consequence, the spotlight of control community turns more and more to data-driven approaches. Data-enabled Predictive Control (DeePC) is such an approach. Built on behavioural systems theory, it refers to Willems’ Fundamental Lemma and replaces the parametric state-space model at the core of the Model Predictive Control framework by data trajectories. While benefiting from advantages of the MPC framework, like including safety constraints, DeePC is purely based on measured data and independent of any parametric system model. Unfortunately, the mathematical foundations of DeePC are strongly tied to linear timeinvariant systems. Hence, challenges arise, when applying DeePC to nonlinear systems. Within this thesis, we explore nonlinear extensions of DeePC and derive heuristics for its hyperparameters by means of simulation. Using these insights, we step into the real world and use DeePC for the control of a Menzi Muck M545 12 t walking excavator. It is the first time, that DeePC is successfully applied to a strongly nonlinear real world mechanical system of such dimensions. After serving a proof of concept, we process an experimental study, and investigate DeePC’s performance on the real world machine in terms of tracking accuracy and robustness.