Optimalisatiegebaseerde Iteratief Lerende Controle: Geavanceerde Methoden voor Uitdagende Mechatronische en Robotica Toepassingen

The increasing demand and competitive economy necessitate the continuous development of intelligent mechatronic and robotic systems capable of adapting to changes and learning over time to perform tasks more accurately, efficiently, and safely. This progress has driven optimization and learning features into control algorithms. This research focuses on enhancing the performance of systems where the same task is executed repetitively. Iterative Learning Control (ILC) is a popular method for achieving this goal by adapting the control input, based on the error observed in previous trials. However, traditional ILC algorithms assume linear dynamics and do not account for constraints in the task specification, e.g. actuator saturation and physical limitations of the systems, which are nowadays critical for performance and safety. Additionally, the task optimization of highly nonlinear dynamics, e.g. quadrotors and robotic manipulators, requires the use of accurate nonlinear models. This thesis aims to develop optimization-based nonlinear ILC algorithms that enable challenging applications in mechatronics and robotics, such as improving performance for a fleet of similar systems, vibration-free handling of flexible objects with a robot manipulator, and the use of low-informative measurements, such as from binary sensors, for improving performance of nonlinear systems. Several challenges are addressed: (1) theoretical analysis of learning control algorithms, (2) learning from multiple systems, (3) combining parametric and nonparametric corrections, (4) dealing with the complex interplay between learning and control, (5) handling partially observable dynamics and quantized measurements, and (6) managing system uncertainty. Firstly, this thesis proposes a Multi-System ILC (MSILC) algorithm. The approach is theoretically analyzed and validated on a numerical example of a fleet of positioning stages, demonstrating that MSILC can improve the fleet's performance compared to traditional ILC by offering superior robustness against non-repetitive disturbances. Secondly, an optimization-based nonlinear ILC is proposed for handling a flexible object with a robot manipulator. The approach combines parametric and nonparametric model corrections, requiring only interoceptive sensors of the robot, and compensating for the unmodelled residual dynamics of the flexible object. The approach is validated on a Franka Emika Panda robot, showing that the proposed ILC algorithm can achieve vibration-free handling of flexible objects. Finally, an Iterative Learning Dual Control (ILDC) approach is proposed for nonlinear systems with binary measurements. The algorithm includes a set-membership estimation algorithm for the parameter uncertainty, a control-relevant estimation, and a dual control strategy balancing exploration (reducing parameter uncertainty) and exploitation (enhancing task performance). The approach is experimentally validated on a cart-pendulum setup with quantized measurement of the load position, demonstrating that the proposed ILDC algorithm can achieve desired performance in several scenarios, even when the quantized sensor is placed outside the load position range covered by the task.