Autonomous Iterative Learning Control: Efficient Real-World Learning of Reference Tracking despite Unknown Dynamics

In order to be capable of solving tasks in the real world, mechatronic and robotic systems have to precisely perform desired motions. Conventionally, this is achieved by model-based control which enables reference tracking of dynamical systems and, hence, the execution of desired motions. However, model-based control comes at the cost of requiring tremendous expertise and manual design efforts. In this thesis, the class of Autonomous Iterative Learning Control (AILC) is introduced, and three learning methods, respectively frameworks, are proposed that belong to the class of AILC and are capable of self-reliantly learning to solve reference tracking tasks. AILC methods incorporate the concept of iteratively updating a feedforward input trajectory to track a desired reference trajectory from Iterative Learning Control (ILC). On top, AILC methods iteratively use experimental data to learn models of the system dynamics online and, hence, overcome the requirement of prior model information. Furthermore, AILC self-reliantly determines necessary learning parameters based on data and the learned models to overcome the need of manual parameter tuning. In summary, AILC enables learning of input trajectories to solve reference tracking tasks in real-world systems without requiring prior model information or manual parameter tuning. This is in stark contrast to the state of the art, in which learning control methods either require prior model information or manual tuning of learning parameters. Lastly, two modular extensions are proposed, which enable the application of the proposed learning methods to systems with Multi-Input/Multi-Output (MIMO) dynamics and output constraints. A particular focus of this thesis lies on the extensive validation of the proposed learning methods using a variety of reference tracking tasks and systems – both in simulation and real-world experiments. One of the main achievements of this thesis is that one single method learns to solve a total of nine different reference tracking tasks across three different real-world systems without requiring prior model information or manual tuning, and while only requiring roughly 5-10 trials of learning per task. This, again, is in stark contrast to the state of the art, in which learning control methods are primarily validated in simulations, and the methods, which are validated by real-world experiments, are typically only validated on a single system and task to which the methods are manually adjusted. In summary, the proposed AILC methods enable mechatronic and robotic systems to self-reliantly learn to solve reference tracking tasks in the real world and hence perform desired motions without requiring the tremendous expertise and manual design efforts that are typically required by model-based control.