Vehicle motion control in handling limits, including drifting control, involves highly nonlinear system dynamics and uncertainties. Conventional dynamic-model-based approaches often require extensive parameter tuning, since fixed parameters are difficult to handle dynamic environmental changes. To address this, a Model Predictive Control (MPC) framework with parameter self-adaptation via reinforcement learning (RL) is proposed. The RL agent can autonomously adjust MPC controller parameters based on its learned experiences, and is capable of online learning during closed-loop control. This framework is first validated on the high-fidelity simulation platform CarSim, showing that the algorithm can achieve stable drifting under various conditions and effectively adapt to dynamic environmental changes. Further, real-vehicle testing is conducted on a full-size B-class electric car with rear-wheel-drive and steer-by-wire systems. According to the published literature, this is the first time that RL methods are successfully trained and deployed on real-vehicle drifting tasks, outperforming the conventional MPC-based algorithm.

Precision Immobilization Technique (PIT) is a potentially effective intervention maneuver for emergency out-of-control vehicle, but its automation is challenged by highly nonlinear collision dynamics, strict safety constraints, and real-time computation requirements. This work presents a PIT-oriented neural optimal-control framework built around PicoPINN (Planning-Informed Compact Physics-Informed Neural Network), a compact physics-informed surrogate obtained through knowledge distillation, hierarchical parameter clustering, and relation-matrix-based parameter reconstruction. A hierarchical neural-OCP (Optimal Control Problem) architecture is then developed, in which an upper virtual decision layer generates PIT decision packages under scenario constraints and a lower coupled-MPC (Model Predictive Control) layer executes interaction-aware control. To evaluate the framework, we construct a PIT Scenario Dataset and conduct surrogate-model comparison, planning-structure ablation, and multi-fidelity assessment from simulation to scaled by-wire vehicle tests. In simulation, adding the upper planning layer improves PIT success rate from 63.8

With the rapid advancement of automated driving, motion sickness (MS) has gathered significant attention from both academia and industry. Advanced sensing, decision-making, and control systems in autonomous vehicles offer promising opportunities to mitigate MS in next-generation transportation. Traditionally, objective metrics such as Motion Sickness Dose Value (MSDV) and Motion Sickness Incidence (MSI) are derived directly from vehicle motion states to quantify MS severity. However, these metrics often overlook the filtering or amplification effects of human body dynamics on the motion stimuli experienced in occupant heads. To address this, we designed two motion planning algorithms to generate vehicle trajectories: one optimized solely based on vehicle motion (control group) and another incorporating occupant head motion dynamics (experimental group). Real-road experiments with 23 participants, using a within-subject design, were conducted on an automated vehicle. Results demonstrated that trajectories with 'head-motion-first' concept can significantly reduce MSDV, with subjective assessments via the Misery Scale (MISC) showing notable reductions in MS severity. This study represents one of the few occupant-in-the-loop on-road validations of MS mitigation through automated driving, confirming the effectiveness of incorporating head motion dynamics into MS-oriented trajectory planning.

Personalized autonomous driving systems are critical to improving user acceptance and trust. However, existing personalization methods mainly rely on imitating manual driving behaviors, which often fail to consider the significant shift in user preferences when transitioning from an active “driver” role to a passive “supervisor” role. Furthermore, traditional learning-based methods face challenges in computational scalability, impeding efficient online adaptation to real-time user feedback. To address these issues, this paper proposes a user-feedback-driven personalized autonomous driving framework that integrates online learning and optimal control. First, a Sparse Incremental Gaussian Process Regression (SI-GPR) framework with a tailored sample management mechanism is introduced. Distinct from conventional modular integration, our framework is uniquely architected to capture the non-stationary transition of user preferences during the role shift from active “driver” to passive “supervisor.” SI-GPR effectively models evolving psychological expectations by dynamically updating a sparse inducing set that prioritizes recent human feedback. Crucially, this mechanism maintains constant computational complexity. Second, the learned personalized parameters are embedded into a Constrained Optimal Control Problem (OCP) for trajectory planning, which dynamically balances safety, comfort, and tracking efficiency. Finally, the proposed framework is validated using a Human-in-the-Loop (HIL) simulation platform. Experimental results supported by subjective ratings and physiological indicators such as heart rate variability demonstrate that the proposed method significantly reduces user interventions and mental workload compared with baseline methods. This work provides a closed-loop solution for aligning autonomous driving behavior with the evolving user expectations in the supervisor role.

Motion sickness (MS) is becoming an increasingly important issue in autonomous driving, as more passengers engage in non-driving-related activities and become more sensitive to ride comfort during automated travel. This paper presents a learning-enhanced nonlinear optimal control framework for personalized MS mitigation in autonomous vehicles. The problem is formulated as a constrained finite-horizon optimal control problem that directly minimizes head-level Motion Sickness Dose Value (MSDV) under vehicle-dynamics and safety constraints. To represent the nonlinear and passenger-specific mapping from vehicle motion to head motion, a personalized 6-degree-of-freedom Vehicle-to-Head Surrogate Model (VHSM) based on a Discrete Cosine Transform (DCT)-enhanced Transformer is embedded into the optimization framework as a differentiable and symbolically deployable surrogate mapping. By extracting frequency-domain features in the MS-sensitive band, the model preserves the dominant physiological motion components while remaining robust to changes in discrete planning step size, enabling flexible alignment between data-driven prediction and downstream nonlinear optimization. Comprehensive simulations across nine driving scenarios show that the proposed Head-Based (HB) planning consistently outperforms conventional Vehicle-Based (VB) planning. Real-vehicle closed-loop experiments with 23 participants further demonstrate that, relative to VB planning, the HB framework reduces the measured head-level MSDV (MSDVh) by 7.6% and Motion Sickness Incidence (MSI) by 7.2% during 16-minute sessions. These results support differentiable surrogate-enabled optimal control as a practical path toward human-centered comfort control in autonomous driving.

This paper proposes a novel decision-making framework for automated vehicles (AVs). The framework integrates traffic situation assessment into the decision process to emulate experienced human drivers. It enables the AV to make forward-looking decisions based on macroscopic traffic flow states, guiding the AV toward regions that are safer, more efficient, and more comfortable. To this end, a three-dimensional (3D) traffic-situation quantification model is developed. The model considers static and dynamic flow complexity, vehicle interaction intensity, and traffic stream motion. A machine-learning-based predictor is then introduced to capture the spatiotemporal evolution of traffic situations. A Level- game framework is further used for situation-aware interaction modeling and behavior selection. The selected decisions are refined by an optimal-control-based trajectory optimizer with kinematic and collision-avoidance constraints. Simulation results in smooth and congested traffic scenarios show that the proposed method improves efficiency and comfort over human driving while maintaining safety.

Heavy-duty truck electrification is accelerating as a critical pathway for freight decarbonization, propelled by maturing sensor systems, ubiquitous vehicular networking, and sophisticated chassis actuation. Unlike electric cars and conventional diesel trucks, battery-electric heavy trucks face intensified trade-offs between load capacity and ride comfort, stemming from the substantial mass of onboard batteries. This paper proposes a Model Predictive Control (MPC)-based coordinated control strategy for heavy truck's full-air suspension systems, addressing the dynamic coupling between chassis suspension and cab mount. A multi-layer "suspension-frame-mount-cab" coupled model is established from real vehicle tests, and an MPC controller is designed to enhance cab comfort through real-time stiffness coordination of both subsystems. Results show 66.56% maximum and 5.45% average improvement in dynamic response over conventional systems, offering a practical framework for intelligent suspension coordination in heavy trucks.

Accurate prediction of vehicle collision dynamics is crucial for advanced safety systems and post-impact control applications, yet existing methods face inherent trade-offs among computational efficiency, prediction accuracy, and data requirements. This paper proposes a dual Physics-Informed Neural Network framework addressing these challenges through two complementary networks. The first network integrates Gaussian Mixture Models with PINN architecture to learn impact force distributions from finite element analysis data while enforcing momentum conservation and energy consistency constraints. The second network employs an adaptive PINN with dynamic constraint weighting to predict post-collision vehicle dynamics, featuring an adaptive physics guard layer that prevents unrealistic predictions whil e preserving data-driven learning capabilities. The framework incorporates uncertainty quantification through time-varying parameters and enables rapid adaptation via fine-tuning strategies. Validation demonstrates significant improvements: the impact force model achieves relative errors below 15.0

Safety validation of autonomous driving (AD) systems remains a critical challenge before large-scale AD deployment on the public road. To fulfill this, the corner case testing plays an important role, especially for those AD function tests involving complex traffic interactions. Traditional testing case generation methods are plagued by fundamental limitations in identifying rare failure events and accurately estimating failure probabilities, which further affects both the testing effectiveness and efficiency. Autonomous emergency braking (AEB) systems, as a key feature guaranteeing the safety of autonomous driving, need to handle the complex interactions with various road users, including vulnerable road users (VRU), e.g. pedestrians and non-motorized vehicles (NMV). Taking AEB test case generation as the application, this paper proposes a Failure Index-Guided Adaptive Importance Sampling (FIGAIS) method designed for corner case test case generation that enables accurate failure probability estimation in AEB system scenarios with NMV interactions. The method introduces an innovative failure density concept that combines local empirical estimation with physically meaningful failure indices through an adaptive weighting mechanism to achieve robust generation of test cases that reflect accurate failure probabilities. A gradientbased optimization algorithm with repulsion mechanism identifies multiple dominating points in the parameter space and constructs a Gaussian Mixture Model (GMM) as the importance sampling proposal distribution. Experimental validation in complex six-dimensional AEB-VRU interaction scenarios demonstrates that FIG-AIS achieves an improvement in test case generation for accurate failure probability estimation and reduced relative error compared to standard Monte Carlo and cross-entropy methods while improving the identification capability for rare failure regions. The method successfully identifies multimodal failure behaviors that traditional testing tends to overlook, providing an efficient and systematic solution for autonomous driving system safety validation.

Aiming at the trajectory planning problem of autonomous vehicles, a spatio-temporal joint planning framework considering both multi-vehicle interactions through a game theoretic approach and asymmetric risk field theory was proposed in this article. Through game theoretic forward propagation, the predicted future trajectory of the surrounding vehicle is acquired and coupled into the framework of the ego vehicle decision-making and planning, so that the ego vehicle trajectory planning considering multi-vehicle interaction can be realized. The trajectory is derived from a spatio-temporal planning approach to integrate the velocity planning and path planning and the safety of the trajectory is guaranteed. Furthermore, the trajectory points generated by forward propagation can effectively consider the asymmetric risk field generated by surrounding vehicles and integrate it into the solution of the numerical optimization problem, comprehensively considering the impact of different types of surrounding vehicles, together with their states and other factors, so that the calculated route is safer and in line with human decision-making. We further leverage the prospect theory to mimic the behavior of real human drivers, considering irrational factors of human decision processes. The simulation results show that in the dense traffic flow with frequent interactions with surrounding vehicles, the autonomous ego vehicle can reasonably change lanes and achieve efficient and safe driving.

Driving at intersections often differs from regular car-following scenarios, particularly in terms of start-up speed uncertainties. Extracting useful information from these uncertainties is crucial for improving traffic and energy efficiency, ultimately achieving eco-driving goals for automated vehicles. However, traditional car-following models are inadequate for describing intersection driving behaviors, while current eco-driving strategies often lack a thorough consideration of intersection traffic uncertainties. To this end, a new eco-driving strategy based on stochastic model predictive control (SMPC) is proposed, specifically addressing uncertainties in the starting-up behavior of leading vehicle (LV). To begin, a naturalistic driving dataset is constructed for a queuing scenario at a signalized intersection, enabling the establishment of probabilistic models for LV starting-up behavior. Next, an SMPC-based motion planning algorithm is introduced to facilitate optimal control of the ego vehicle (EV), directly incorporating the LV’s uncertainties. Specifically, a phantom LV model is introduced to depict LV’s starting-up probability in its initial preparing stage, and then in the accelerating stage a gaussian process regression (GPR) approach is employed for more accurate prediction of LV acceleration. Then the algorithm’s effectiveness is validated through stochastic and playback simulations in SUMO and Matlab. The playback simulations using conditions from naturalistic driving demonstrate that the algorithm significantly enhances energy efficiency, achieving savings of 21% and 37% compared to electric and fueled vehicles in real-world scenarios, respectively. Surprisingly, despite expectations, the algorithm can also improve traffic efficiency by an average of 17.9% and 13.6% for electric and fueled vehicles, respectively. Finally, the road experiment validates that our algorithm can achieve total savings of 15.8% energy in our experimental electric vehicle compared to real drivers. This algorithm has the potential to serve as an effective eco-driving solution for automated vehicle longitudinal planning and control in urban settings.

Motion sickness (MS) has long been a common complaint in road transportation. However, in the era of driving automation, MS has become an increasingly significant issue. The future intelligent vehicle is envisioned as a mobile space for work or entertainment, but unfortunately passengers' engagement in non-driving tasks may exacerbate MS. Finding effective MS countermeasures is crucial to ensure a pleasant passenger experience. Nevertheless, due to the complex mechanism of MS, there are numerous challenges in mitigating it, hindering the development of practical countermeasures. To address this, we first review two prevalent theories explaining the mechanism of MS. Subsequently, this paper provides a summary of current subjective and objective approaches for quantifying motion sickness levels. Then, it surveys existing methods for alleviating MS, including passenger adjustment, intelligent vehicle solutions, and motion cues of various modalities. Furthermore, we outline the limitations and remaining challenges of current research and highlight novel opportunities in the context of intelligent vehicles. Finally, we propose an integrated framework for alleviating MS. The findings of this review will enhance our understanding of carsickness and offer valuable insights for future research and practice in MS mitigation within modern vehicles.

This paper presents a comprehensive Pedestrian-Vehicle Interaction Behavior Model (PVIBM) grounded in the fundamental logic of human behavior, decomposing pedestrian actions into three sequential processes: perception, decision-making, and action. Each subprocess is modeled separately to achieve a holistic understanding of pedestrian-vehicle interactions. To infer vehicle intentions, a role-immersing experiment is designed to collect relevant data, which is then analyzed using machine learning techniques. Pedestrian actions are extensively modeled by adapting the dual accumulator model within a game-theoretic framework, incorporating a specifically designed utility function for pedestrian-vehicle interactions. Genetic algorithms are employed for parameter identification of the proposed PVIBM using the inD dataset [1], yielding a reasonable parameter distribution and high model accuracy. The results from case studies and simulations demonstrate that PVIBM not only offers interpretability but also produces pedestrian decision-making outcomes and speed distributions closely resembling real pedestrian behavior.

As a critical driving scenario, emergency collision avoidance (CA) is challenging and necessary for both human drivers and automated vehicles. Compared to passenger cars, the complexity of tractor-trailer vehicle dynamics makes it more difficult to implement CA features, especially when designing the motion planning module. Using traditional planning methods, it is difficult to balance optimality, safety and real-time performances. To address these needs, this study proposes an optimal-control-problem (OCP) based framework of motion planning for tractor-trailer vehicles, with specific focuses on the fast solution of OCP that considers nonlinear vehicle model, state constraints and roll-over stability limits. Inspired by expert drivers' experience in planning within a previewed distance and then keeping updating, the original time-domain OCP with free terminal time is transformed into a space-domain OCP with fixed length, which can achieve fast, robust and forward-looking motion planning. Simulations and scaled vehicle experiments show that the planning algorithm performs safely and efficiently under different speed and road adhesion conditions.

Due to highly dynamic interactions between vehicles, ramp merging decision is extraordinarily challenging in dense traffic. To deal with the non-convexity and inter-coupling of the decision-making problem, a top-level behavioural planner is proposed, which considers both the overall traffic situation and the individual characteristics of other interacting drivers. To ensure the implementability and efficiency of planning, a bottom-level motion planner is further designed with the guide of the top-level behavioural planner. The validation simulation using a naturalistic driving dataset shows that the proposed planning algorithm can achieve a success rate of 97.65% and has similar gap selection decision as human drivers. Then to track the planned vehicle motion, a nonlinear model predictive controller considering actuator delay and lag characteristics is proposed. Finally, the proposed planning and control modules are deployed in a turbocharged test vehicle, with satisfactory vehicle lateral and speed tracking errors, which validate the implementability of the proposed ramp merging decision algorithms.

Driving in scenarios with occlusion is challenging but common in daily traffic, especially in urban and rural areas. To handle the potential interaction between the ego vehicle and pedestrian that possibly exists but is occluded by front vehicle, a stochastic model predictive control (SMPC)-based motion planning algorithm is proposed in this study. Firstly, a naturalistic driving dataset of vehicle-pedestrian interaction is established, based on which it is found that in the case of pedestrians passing or not, there are significant differences in front vehicle driving behavior. Then, a probability estimation approach for the presence of pedestrians in the occluded area is designed, which can achieve 91.9% accuracy in the naturalistic driving dataset. A phantom pedestrian model is established to quantify the uncertainty in the occluded area, which is further used to construct the chance constraint of the SMPC planning problem. Finally, a naturalistic driving data based simulation and a pedestrian-driver-in-the-loop experiment are carried out to validate the proposed algorithm. Both simulation and experiments show that our algorithm can effectively utilize the perceived information to speculate pedestrian presence beyond sensing range, thereby enabling proactive decisions to achieve safety, comfort and traffic efficiency in vehicle-pedestrian interactions. The proposed framework may find applications in interaction planning problems with uncertainty challenges.

The prevalence of motion sickness among passengers using personal electronic devices, such as smartphones, during vehicle journeys has become a growing concern. This issue is expected to intensify with the increasing adoption of assistant or automated driving functions, which may lead to non-driving tasks (NDT) being performed by all on-board passengers, including the user in the "driver" seat during conditionally or fully automated driving modes. This trend presents challenges related to motion sickness, particularly in terms of specific performance requirements for non-driving tasks. In response to the need to alleviate passenger motion sickness, we have developed an easy-to-understand animation cue app that can be conveniently implemented on smartphones. The motion cue conveys information about vehicle accelerations, including their directions and magnitudes, using the metaphors of traffic signal colors and backward-moving lane lines, either in straight or curved lane driving. Following several rounds of improvements based on moving-base simulator and real car experiments, finally a successful cue design was found, which could significantly alleviate motion sickness of passengers while engaging in NDT, with minimal impact on their NDT performances. However, the study also revealed limitations to the sickness-alleviating capability of our motion cue design, including potential lack of universal acceptance among different users and reduced effectiveness in severely uncomfortable driving conditions. This work may provide valuable insights for further visual cue improvements that can contribute in future carsickness-proof vehicles.

For autonomous driving, it is important to develop safe and efficient decision algorithms to handle multi-vehicle interactions. Game theory is suitable to manage the interactive driving decision modelling, however, common approaches of multi-player game formulation is computationally complex for dynamic and intense interactions. The main contributions of this work are two-fold: 1) a global sorting-local gaming framework, namely GLOSO-LOGA, is proposed to solve the intersection interaction problem for autonomous driving, which can comprehensively consider the advantages of multi-vehicle collaboration and single-vehicle intelligence approaches; 2) an interaction disturbance function is used to quantify the impact of indirect interactions on ego vehicle. To validate the algorithm performances, corner case simulations and human-in-the-loop simulator experiments are carried out, in which a four-armed intersection scenario with various urgent and challenging interaction conditions is used. Results show that compared to the traditional approach that decomposes a multi-vehicle game into multiple two-vehicle games, the proposed algorithm can improve both safety and traffic efficiency in intensively interactive driving scenarios even in complex and urgent cases. It may be potentially applied in handling autonomous vehicle's dynamic interactions with multiple road users.

The personalization or customization of automated driving algorithms has attracted great attention because of its potential in enhancing trust and acceptance. This is true for both Advanced Driving Assistance Systems (ADAS) and higher-level functions such as automated driving. Currently, the common practice of personalized algorithm mainly focuses on reproducing users' naturalistic driving operations while still treating them as drivers, even if they are not necessarily in the control loop of driving tasks. Taking a new perspective from the user role transition between manual and automated driving modes, this work focuses on users' preferences in selecting potential interacting agent(s), i.e. the process of Zone-of-Interaction (ZOI) prioritization, based on their visual attention characteristics. Firstly, driving simulator experiments are conducted, and subjects' attention allocation preferences in both automated and manual driving modes are analyzed. Considering personal preferences in time and spatial distributions of attention, Markov Decision Processes (MDPs) are then adopted to model the personalized ZOI prioritization in automated driving across different scenarios. Simulation results based on random settings and real user data show that the proposed model can decide a ranking of potential interacting agents that is consistent with users' personal expectations. Further application in personalized automated driving is conducted, which indicates that the personalized ZOI prioritization model can significantly reduce users' mental workload and enhance user acceptance. Our work can hopefully be applied in developing personalized planning algorithm of automated driving, especially for complex scenarios involving multiple traffic agents.