Robust Policy Optimization

• Applicant: Professor Dr.-Ing. Christian Ebenbauer
• Subject Area: Automation, Mechatronics, Control Systems, Intelligent Technical Systems, Robotics
• Term: since 2025
• Project identifier: Deutsche Forschungsgemeinschaft (DFG) - Project number 550848323

Trajectory optimization is a fundamental task in control engineering. It comprises the generation of reference trajectories and tracking controllers. There are many different fields of application for trajectory optimization, such as robotics as well as aerospace and automotive engineering. A common approach to trajectory optimization is the classical two-degree-of-freedom control architecture. It consists of a feedforward and a feedback term. The role of the feedforward term is to achieve the nominal control objective, such as reaching a desired target state or satisfying control constraints. The feedback term serves to counteract deviations from the reference trajectory provided by the feedforward term. These deviations naturally occur due to disturbances or modeling errors. In this sense, the latter term improves the robustness of the control architecture. Robustness plays a key role in many trajectory optimization problems. This is especially true for trajectory optimization problems in environments with high uncertainties and in applications in which specific safety standards must be met; e.g., in aerospace and automated driving applications. Despite various research efforts, there are important open problems regarding robustness. In particular, powerful strategies to optimize robust stability and robust performance of trajectory generation and tracking architectures for various classes of uncertainties are still lacking. To address these challenges of robustness, we aim to develop a novel methodology, which will be based on the ideas and principles of robust control. Our research objectives are not limited to trajectory optimization, but, more generally, they address the problem of optimizing control policies (control strategies) of affine linear form. This is motivated by the classical two-degree-of-freedom architecture. Affine linear policies also appear in the synthesis of neural networks, which is a another potential field of application for our new methodology. The theory of robust control makes it possible to describe large classes of uncertainties and to treat linearization errors. We will use this theory to develop algorithms that synthesize policies, which guarantee robust performance and robust constraint satisfaction.Most of the existing methods are too computationally expensive for online implementations. For this reason, we will also focus on computational aspects of robust policy optimization. For example, we will employ ideas from differential dynamic programming to enable fast computations of local feedback solutions. To further enhance computational efficiency, we will also investigate structure-exploiting algorithms. The overall goal of this research project is to develop an online-implementable methodology for robust policy optimization. Our new approach is intended to be applicable to general nonlinear control systems with different types of uncertainties and constraints.

• DFG Programme: Research Grants
• Co-Investigator: Professor Dr. Carsten Scherer