Rolf Findeisen

TU Darmstadt, Control and Cyber Physical Systems Laboratory

Tuesday, February 25, 2025, 11:00 - 12:15

SR 01-012

Abstract:

Machine learning has shown great potential in improving control performance through adaptation. model identification, controller learning, and data-driven optimization. However, ensuring safety, constraint satisfaction, and long-term optimality while making short-term, computationally efficient decisions remains a major challenge, particularly in real-world applications. This talk explores how machine learning can be integrated with control to enable safe and efficient decision-making in repetitive tasks, such as human support via exoskeletons, robotic operations, and autonomous driving. By leveraging Gaussian processes, neural networks, and Bayesian optimization, we learn control-relevant parameters—such as cost functions, constraints, or model components—while ensuring stability, constraint satisfaction, efficient implementation, and safety.

To achieve this, we employ a two-level optimization framework that enables safe exploration in real-world experiments while efficiently transferring knowledge from simulations. We present methods for Bayesian safe optimization, domain adaptation, and leveraging incomplete experimental data, enabling robust learning in practical settings. The results, validated through simulations and experiments, demonstrate how learning-based approaches can enhance control performance without compromising safety. Ultimately, this work paves the way for real-time learning and adaptation in control systems, ensuring both long-term optimality and operational reliability in complex engineering applications.

CV:

Rolf Findeisen’s research focuses on model predictive control, nonlinear optimization-based control, and learning-based control methods. He is a Professor at TU Darmstadt, where he leads the Control and Cyber-Physical Systems Laboratory. His work integrates machine learning with predictive control, ensuring safety and long-term optimality in decision-making for complex systems. His applications span autonomous systems, robotics, process control, medical engineering, and cyber-physical systems. He has contributed significantly to the development of optimization-based control strategies.