In this paper, a combined control structure is presented integrating a classical Linear Quadratic Regulator (LQR) with a machine learning-based nonlinear compensator implemented through Kolmogorov-Arnold Networks (KANs). The main goal of the presented method is to ensure stability guarantees and computational efficiency, while the control performance level is increased. Firstly, a Reinforcement Learning (RL) framework is used to learn an optimal nonlinear policy over the operational range of the nonlinear system. Then, the trained agent is used as a teacher network for supervised knowledge distillation into a KAN. Finally, a stability condition is derived using the maximum Lipschitz constant of the KAN to guarantee the stability of the closed-loop LQR-KAN system. The results show that the KAN-based control structure achieves performance levels nearly the same as the RL policy, while the stability analysis is significantly simplified and the resulting neural network is less complex than the original teacher network. The whole approach is demonstrated through a double inverted pendulum problem.

Among modern traffic management solutions, artificial intelligence-driven approaches are becoming increasingly prevalent, addressing complex operational tasks through datadriven methods rather than traditional physical modeling of system dynamics. Large Language Models play an increasingly significant role in this domain, as they demonstrate substantial potential for zero-shot performance in reducing green-house gas emissions and minimizing delays within complex transport networks. This paper presents a novel methodology where, by establishing a robust semantic connection between real-time sensor data and the task description defined in the prompt, the deployed Large Language Model performs comprehensive multiobjective optimization. This optimization process simultaneously accounts for critical sustainability metrics-specifically green-house gas emissions and environmental noise pollution-and factors ensuring traffic flow continuity, with a primary focus on minimized waiting times and the highest possible realized average speeds. The research demonstrates that the resource-intensive training process can be bypassed; through effective prompting, the autonomous agent evaluates the current state of the network to determine optimal Variable Speed Limits on the highway, relying solely on sector-specific average speeds. Ultimately, this study introduces a validated solution within a simulation environment that provides tangible mutual benefits for both residents living in the vicinity of highways and the commuters traveling on them.

Motion planning of unmanned aerial vehicles (UAVs) for coordinating with ground vehicles (GVs) has a growing impact on industrial and logistic systems, e.g., in the delivery and supervision process. The actual challenge of the coordination is to provide methods for achieving safe and time-efficient motion profile for the UAVs. Moreover, a high variety of UAVs requires a modular control structure with limited real-time computational effort. This paper presents a control design method for an individually controlled UAV to adapt its motion for avoiding moving GVs. The velocity planning layer contains two control elements: reinforcement learning (RL) based controller and quadratic optimization based supervisor. The RL-based controller aims to reduce the mission time of the UAV and similarly, the supervisor provides guarantees on its collision-free motion. The performance of the proposed control algorithm is evaluated through comparative simulations and real-life implementations. Moreover, the operation is also analyzed through simulation scenarios involving multiple UAVs and GVs for determining achievable overall traveled distance performance within the working process.

The aim of this study is to propose a reinforcement-learning-based synthesis method for the integrated design of a robust ℋ∞ control and a deep neural network. This integration improves closed-loop performance through Proximal Policy Optimization (PPO), i.e. the ℋ∞ controller and the deep neural network are trained within the same optimization process; this joint training constitutes the main novelty of the paper. The resulting control system provides performance guarantees against uncertainties which are taken into account both in the ℋ∞ design process and in the training environment. The effectiveness of the integrated design method is illustrated through the coordinated control of simultaneously moving robot vehicles, where reduced mission time and collision avoidance are achieved even under significant uncertainties.

In the era of autonomous vehicles, state estimation is crucial for planning and control. In this paper, improvements for the filtering of wheel odometry & GNSS measurements are proposed to increase the accuracy in the challenging city driving localization task. Although wheel encoder-based estimation is a robust and cost-effective localization method in passenger vehicles, its accuracy is limited by parameter uncertainty; therefore, the proposed method includes online calibration of model parameters. The avoidance of pose estimation divergence can be guaranteed with accurate GNSS observations, thus the estimation of the current measurement noise is also considered. This information is fed to a neural network to predict optimal covariances for a Kalman-filter. The effectiveness of the proposed method through experimental measurement is illustrated, which is saved in urban driving, where the standard deviation of the GNSS is in the meter range. Furthermore, the proposed method can operate with any kind of noisy pose measurements, e.g. from vision even indoor, thus it has a wide range of practical applications in the field of autonomous robotics.

This paper addresses the design of a Cooperative Adaptive Cruise Control (CACC) controller for a heterogeneous vehicle platoon, considering variations in driveline dynamics. The proposed control architecture relies on a Predecessor-Follower communication topology, and a spacing policy where the time headway depends on vehicle speed is adopted. The control problem is formulated within the ℋ∞ linear parameter-varying (LPV) framework and solved using both grid-based and polytopic LPV design methods. The effectiveness of the designed controller is demonstrated through frequency-domain analysis and time-domain simulations.

This paper introduces a novel method for the enhancement of automated vehicle safety and efficiency during critical manoeuvres. The fundamental of the presented method is the observer design architecture, in which lateral dynamic states of the vehicle are evaluated. The novel observer consists of both model-based and machine-learning-based methods to ensure the selected design performances, such as efficient trajectory tracking and safety evaluation of the autonomous vehicle. In contrast to the already introduced and applied stability index-based methods, the proposed safety evaluation process is able detect stability loss and performance degradation of the autonomous vehicle. In the proposed observer-based safety evaluation method, stability and performance loss detection is based on the comparison of model-based and learning-based state observation. The main novelty of the paper is the design of the reinforcement learning (RL) based observer in a guaranteed structure that results in small observation error even under nonlinear vehicle dynamics. Furthermore, a lateral safety index is defined based on the value of the improvement vector representing the addition to the model-based estimation. By this means, with the proposed safety evaluation method both safety and performance loss hazards can be identified simultaneously.

This paper introduces a reinforcement learning (RL)-based framework for adaptively combining four different control strategies in response to varying operational conditions. Starting from a nominal model of the actual system, several feedback controllers are developed, each offering distinct performance benefits under different circumstances. The RL algorithm dynamically determines and mixes the outputs of the controller within specific operating ranges. Four control approaches are considered: Linear Parameter Varying (LPV), Ultra-local Model-based (ULM), Linear Quadratic Regulator (LQR), and a kinematic model-based method. The proposed solution is validated through different test scenarios using the high-fidelity vehicle simulation platform, CarMaker.

This paper presents a method for improving the performance level of the robust H-infinity controllers. The improvement is achieved in an additive form, which contains a learning-based agent. The contribution of the presented method is that the design methods of the H-infinity controller and of the learning-based agent are formed in a joint optimization. This results in the iterative design of the controllers within a reinforcement learning algorithm. The developed design method is applied to an autonomous vehicle control problem for lap time minimization. The presented simulation-based analysis shows that the proposed method can provide improved performance level, compared to the conventional H-infinity control without extension or to H-infinity control with extension but without joint optimization. Copyright (c) 2025 The Authors. This is an open access article under the CC BY-NC-ND license (https://creativecommons.org/licenses/by-nc-nd/4.0/)

This research focuses on controlling the motion trajectory of autonomous vehicles by using a combination of two high-performance control methods: Linear Parameter Varying (LPV) and Reinforcement Learning (RL). First, a single-track motion model is researched and developed with coordinate systems to determine the car's motion trajectory through signals from GPS. Then, the LPV control method is used to design a controller to control the car's motion trajectory. Reinforcement learning method with detailed training procedures is used to combine with the advantages of LPV controller. Finally, the simulation results are evaluated in the time domain through the use of specialized CarSim software, which clearly demonstrates the superiority of the research method.

This paper presents a safe control framework for autonomous vehicle systems with high-performance data-based elements, e.g., a reinforcement learning-based agent. The controller is augmented with a classic robust controller and a supervisor to achieve the highest performance while providing stable and robust motion control even in critical situations. The supervisor prioritizes the control signal of the RL-agent, however, it is able to switch to the robust controller in potentially unsafe situations. The decision is based on a model-based predictive algorithm that uses the model of the vehicle to predict future motion. In addition, the predictive layer is able to detect if the vehicle is about to leave the road or crash with a nearby object is imminent by solving a nonlinear constrained optimization problem. The effectiveness of the proposed control method using a simulated scenario is illustrated.

This study concerns a dynamic vertical quarter model of a passenger bus using an innovative air suspension system. A PID controller is designed to adjust the air spring element’s stiffness through pressure modulation. The air suspension system on the vehicle model is compared to traditional leaf spring with alterations in stiffness. The investigation reveals compelling results, showcasing the remarkable improvements achieved by the proposed model and controller. Specifically, the observation makes a remarkable 87.1

The goal of the paper is to provide a fairly complete overview on geometric approaches to design detection filters for (q)LPV systems, which can include LTI, LTV and certain input affine nonlinear systems. First, as illustration, the fundamental problem of the residual generation FPRG will be solved based on the concept of LPV invariant subspaces. Then different other detection approaches for filter design will be presented, i.e., inversion based approaches and also null-space based methods will be elaborated. Copyright (c) 2025 The Authors. This is an open access article under the CC BY-NC-ND license (https://creativecommons.org/licenses/by-nc-nd/4.0/)

In this paper, a Linear Parameter-Varying (LPV) control algorithm that integrates an error-based ultra-local model is presented. At each iteration, an ultra-local model is computed, serving as the basis for identifying the scheduling variable of the polytopic system. Analysis of the identified models presents a high correlation among the ultra-local models and key system parameters, which are challenging to measure with high accuracy in practice. To overcome this limitation, the deviation from the nominal model is computed using the error-based ultra-local model. The proposed approach is validated through a lateral control problem for automated vehicles, which is implemented in the vehicle dynamics simulation software, CarMaker. The results highlight the potential of combining the LPV control framework with the error-based ultra-local modeling technique to increase the adaptability and performance of control algorithms in dynamic environments. Copyright (c) 2025 The Authors. This is an open access article under the CC BY-NC-ND license (https://creativecommons.org/licenses/by-nc-nd/4.0/)

This paper presents a robust ℋ∞ control design method that incorporates learning feature. The goal of the work is to improve the performance level of the robust controller, which is achieved through reinforcement learning (RL) process. The contribution of the presented method is that the design of the ℋ∞ controller and the RL-based controller is structured in a joint optimization. This leads to an iterative design for the controllers. The developed design method is applied to the problem of minimizing lap time in vehicle control design. The presented simulation-based analysis shows that the proposed method is able to provide improved performance level, i.e., reduced lap time, compared to the robust control or the independent design process. The outcome of the joint design is that the lap time achieved with the ℋ∞ controller approaches the lap time achievable with the RL process.

This research delves into optimizing the pressure within air suspension systems installed on tractor semi-trailers, leveraging the Dynamic Road Stress Factor. Initially, two models are proposed: one employing a leaf spring suspension system for all axles, and the other incorporating an air suspension system for the two rear axles of the tractor. The air springs are characterized using the Gensys model, focusing on vertical dynamics. The characteristic curve of an air spring is built based on the excitation frequency. Assessment of tractor semi-trailer performance utilizing air suspension reveals a notable 20% reduction in dynamic forces exerted on the road surface compared to leaf spring systems. Ultimately, optimal air spring pressures are determined through the evaluation of the Dynamic Road Stress Factor for the tractor’s two rear axles. Based on these results, the study has proposed a plan to adjust the pressure of air springs according to different operating conditions.

Deep reinforcement learning is a promising technique that can help create autonomous agents. However, it is still an open problem how one can create a controller with robust operation for real-world automotive systems. The difficulty lies in either sample efficiency for real-world learning or developing a good enough simulator for training. This paper addresses the latter, proposing a method that provides a solution to the sim-to-real gap through domain randomization, learning with disturbances, and observation preprocessing. The method is validated on a small-scale F1TENTH-type test vehicle, that is trained to race autonomously in a fully end-to-end manner. It is demonstrated that the training process results in a policy that can drive the car safely even over the grip limit.

Improving trust in the operation of autonomous vehicle (AV) control systems is an actual challenge for overcoming the trough of disillusionment in the development phase. There are identified critical gaps in the field that motivate the development of new theoretically grounded methods: most of the existing methods can be used for specific systems without generality applicability, and the achieved explainability level is aimed only engineers and specialists. This paper aims to provide a user-centered method in order to be able to explain the operation for general users. The research question is: How is it possible to transform complex control systems to approximating systems with explainable operation? In this paper a decision-tree-based solution is proposed that results in a low-order approximating system in explainable form. The application of the method is illustrated through the example of an AV safety detector. The operation of the complex detector is described through decision-tree-based rules that are visualized to explain the original system.

Improving trust in the operation of intelligent transportation systems (ITSs) is an actual challenge for overcoming the trough of disillusionment in the development phase regarding autonomous vehicles (AVs). There are identified critical gaps in the field that motivate the development of new theoretically grounded methods: most of the existing methods can be used for specific systems without generality in applicability, and the achieved explainability level is aimed only engineers and specialists. This paper aims to provide a method for developing an explainable representation on a specific ITS, such as intersection management with AV. The challenge is to find a transformation method which the explainable representation is resulted in. In this paper a decision-tree-based solution is proposed that results in a low-order approximating system in explainable form. It is presented an optimization method that results in the decision tree through the selection of its parameters, focusing on the selected ITS problem. The achieved rules within the explainable representation are used for supporting the human driving strategy in order to reduce critical interactions between AVs and human-driven vehicles. The effectiveness of the method is illustrated through high number of simulation scenarios. The outcome of the simulation is that the number of critical and risky interactions can be significantly reduced, if the rules from the explainable representation are considered.