Learning-supported and force feedback model predictive control in robotics

Abstract

Autonomous dynamical systems enter our daily lives and homes. They appear in the form of self-driving cars, vacuum cleaning robots, and smart manufacturing cobots. Hence, these systems inevitably interact with humans or their surroundings. Safety and reliable performance are two critical aspects, which autonomous systems must meet. These goals can only be achieved by a tight interconnection between several technologies. This includes, among others, the dynamic control of systems and the ability of these systems to learn from interactions. This thesis aims to interlink these two concepts more closely. In particular, we propose to equip model-based controllers with a measure and awareness of the system surroundings. This goal is tackled from three perspectives: First, model predictive control schemes for direct force control are proposed. This tailored model predictive control formulation regulates the interaction forces between a cobot and its environment. Since these forces quantify the interaction of the robotic system with workpieces, other robots, or humans, direct control of contacts is achieved. We hereby utilise the benefits of model predictive control, which allows handling the nonlinearities of robotic systems, includes prior knowledge in terms of models, and explicitly achieves constraint satisfaction. We show how constraints on the interaction forces allow for improved performance and increased safety. They allow guaranteeing tight contact without exceeding safety-critical force limitations. Consequently, the developed model predictive force controller enable to extend the usage of cobots to applications that might have been considered too delicate or dangerous until now. Second, this thesis examines data- and learning-supported predictive force controllers. We equip the developed model predictive force controllers with machine learning models of the interaction. The contact forces are learned via Gaussian processes, which perform well despite noise in sensor data and allow for incorporating prior knowledge about the interaction. Inspired by force control, we develop a generalised concept. We present how Gaussian processes can be included as system output models in predictive controllers, while we provide closed-loop guarantees. The proposed concepts of output learning for predictive controllers constitute a promising step towards the integration of control engineering and computer science. In particular, we achieve increased autonomy of robotic systems that interact with diverse environments via the proposed algorithms. We enable the transfer of predesigned robot control setups to a large variety of applications and environments with the developed learning-supported force controller. Third, this thesis introduces Gaussian processes reference generators for predictive controllers to address the learning from interactions. Since model predictive controllers rely on predictions of future evolutions, the knowledge of the desired motions over the prediction horizon is required. In many cases, these references are not known a priori but encoded by data and obtained via communication and interaction with other systems. Gaussian processes can be used to model, filter, and predict these signals such that the model predictive controller can proactively and foresightedly steer the system to follow the reference. This thesis proposes constrained learning algorithms for the data-based reference generation by Gaussian processes. The proposed constraints in the training phase of Gaussian processes encode trackability conditions. We guarantee the recursive feasibility of learning-supported model predictive control despite uncertainties. The developed reference learning scheme for predictive control tackles the integration of machine learning and control by including systems theory into the data-based training of Gaussian processes. Hence, we achieve the incorporation of extensive prior knowledge in machine learning to obtain reliab