SmartCorners: Artificial Intelligence Techniques for Vehicle Dynamics Control

17th NOVEMBER 2025 REINFORCEMENT LEARNING-BASED ADAPTIVE CONTROL RL is employed as an adaptive tuning layer for an integrated NMPC that combines comfort enhancement and traction control functionalities. The RL-based weight scheduling dynamically adjusts control priorities in real time based on vehicle states and preview information on road height and friction conditions. Control architecture with RL-based adaptive tuning layer. INTRODUCTION Vehicle dynamics control systems play a crucial role in enhancing driving comfort, safety, and handling performance. While many advanced algorithms demonstrate their potential in simulation environments, their effectiveness often decreases in realworld applications, due to modeling assumptions and the simplification required for real-time implementation. To overcome these challenges, artificial intelligence (AI) techniques can be employed as complementary tools or even as alternatives to traditional controllers . This work explores three AI-based approaches: • Imitation Learning (IL): Enables real-time implementation of complex and computationally expensive controllers. • Reinforcement Learning (RL): Dynamically adjusts the control priorities of integrated controllers with multiple functionalities. • Online Deep Learning (ODL): Continuously learns and adapts to varying driver behaviors in real time. ARTIFICIAL INTELLIGENCE TECHNIQUES FOR VEHICLE DYNAMICS CONTROL Politecnico Di Torino*, AVL Ditest GMBH, AVL Deutschland GMBH, Marelli Suspension Systems Italy SPA, Elaphe Pogonske Tehnologije Doo, Heron Sports, Blueways International BVBA, Armengaud Innovate GMBH, Technische Universitaet Ilmenau, Urban Electric Mobility Initiative (UEMI) GGMBH, AVL List GMBH *Corresponding authors: [email protected] Project website: www.smartcorners.eu IMITATION LEARNING-BASED TORQUE VECTORING CONTROL IL is employed to develop a deep neural network (DNN)-based torque vectoring controller that matches the performance of a nonlinear model predictive controller (NMPC). The process includes: • Control expert optimization: The NMPC is tuned under various driving conditions to improve tracking performance and robustness. • Training database generation: A dataset of some of the most relevant NMPC inputs and outputs is collected from over 300 driving maneuvers. • DNN training: Bayesian optimization is used to identify the optimal DNN architecture. IL process of a NMPC-based torque vectoring controller. More information about SmartCorners here! The scalable SMART CORNER SYSTEM (SCS) with its in-wheel motor, chassis actuators, and vehicle motion control will enable user-centric and purpose-driven design of next-gen EVs, leading to right-sized, flexible, and energy-efficient mobility solutions. This project has received funding from the European Union’s Horizon Europe research and innovation programme under the GA No. 101138110. Views and opinions expressed are however those of the author(s) only and do not necessarily reflect those of the European Union or European Climate, Infrastructure and Environment Executive Agency (CINEA). Neither the European Union nor the granting authority can be held responsible for them. ONLINE DEEP LEARNING-BASED DRIVER BEHAVIOR PREVIEW Driver behavior in cornering can vary widely among individuals: some tend to be more aggressive, while others drive more cautiously. ODL is used to predict driver intent, specifically future steering wheel angles, based on current vehicle states and upcoming road curvature . Moreover, ODL continuously adapts in real time to changes in driver behavior, preventing model accuracy degradation. ODL network example. Lateral error Long short-term memory (LSTM) Heading error Straight line length Discretized road curvature Current steering wheel angle Future steering wheel angles Vehicle speed Phase 2: Training database generation Manoeuvre selector End Start new NMPC EM torques Simulation Optimal NMPC cost function weights Surrogateopt NMPC cost function weights Mean KPIs Phase 1: Control expert optimization DNN Hyperparameters Validation loss Phase 3: Deep neural network (DNN) training Neural network training Bayesian optimization Training database Simulation running toolchain Manoeuvre selector End Start new NMPC EM torques Simulation