In vertical vehicle dynamics control, semi-active dampers are used to enhance ride comfort and road-holding with only minor additional energy expenses. However, a complex control problem arises from the combined effects of (1) the constrained semi-active damper characteristic, (2) the opposing control objectives of improving ride comfort and road-holding, and (3) the additionally coupled vertical dynamic system. This work presents the application of Reinforcement Learning to the vertical dynamics control problem of a real street vehicle to address these issues. We discuss the entire Reinforcement Learning-based controller design process, which started with deriving a sufficiently accurate training model representing the vehicle behavior. The obtained model was then used to train a Reinforcement Learning agent, which offered improved vehicle ride qualities. After that, we verified the trained agent in a full-vehicle simulation setup before the agent was deployed in the real vehicle. Quantitative and qualitative real-world tests highlight the increased performance of the trained agent in comparison to a benchmark controller. Tests on a real-world four-post test rig showed that the trained RL-based controller was able to outperform an offline-optimized benchmark controller on road-like excitations, improving the comfort criterion by about 2.5% and the road-holding criterion by about 2.0% on average.
24 October 2025 in Swedish Aerospace Conference, FT2025
Hybrid simulation models combine physics equations withtrainable components to improve simulation results andperformance. Physics-enhanced neural ordinary differentialequations (PeN-ODE) are a promising type of hybrid modelsthat combine artificial neural networks (NN) with thedifferential equations of a dynamic system. Dynamicalsimulation models are often part of embedded controlalgorithms of cyber-physical systems (CPS); compliance withthe safety and real-time requirements of such embeddedenvironments is, however, challenging.In this work, we propose a workflow to incorporate trainedNNs in Modelica models to form hybrid simulation modelsthat are PeN-ODEs. We thereby focus on the transformationsteps from equation-based trained PeN-ODEs in Modelicatowards causal solutions suited for the embedded domain --up to and including MISRA C:2023 compliance checks andfinal software-in-the-loop (SiL) tests of generatedproduction code in the modeling environment -- for which weleverage eFMI standard compliant tools (Dymola and SoftwareProduction Engineering). It is of particular interest, howthe trained NNs of the hybrid model are implemented. Wepresent two approaches: (1) generation of C code usingexisting Open Neural Network Exchange (ONNX) tooling and(2) pure Modelica code with the tensor-flow represented asmulti-dimensional equations. Both approaches are discussed,highlighting why (2) is, in the long run, a better optiongiven the eFMI technology space.