Jiawei Liu, Xun Gong, Muli Yang, Xingrui Yu, Fen Fang, Xulei Yang, Ivor Tsang, Yunfeng hu, Hong Chen, Qing Guo. Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2026.

Ensuring robust and generalizable autonomous driving requires not only broad scenario coverage but also efficient repair of failure cases, particularly those related to challenging and safety-critical scenarios. However, existing scenario generation and selection methods often lack adaptivity and semantic relevance, limiting their impact on performance improvement. In this paper, we propose SERA, an LLM-powered framework that enables autonomous driving systems to self-evolve by repairing failure cases through targeted scenario recommendation. By analyzing performance logs, SERA identifies failure patterns and dynamically retrieves semantically aligned scenarios from a structured bank. An LLM-based reflection mechanism further refines these recommendations to maximize relevance and diversity. The selected scenarios are used for few-shot fine-tuning, enabling targeted adaptation with minimal data. Experiments on the benchmark show that SERA consistently improves key metrics across multiple autonomous driving baselines, demonstrating its effectiveness and generalizability under safety-critical conditions.

Jiawei Liu, Yanjiao Liu, Xun Gong, Tingting Wang, Hong Chen, Yunfeng Hu. Proceedings of the 2025 Conference of the Nations of the Americas Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 1: Long Papers). 2025.

The joint drift problem may result in the omnidirectional mobile manipulator (OMM) failing to perform its tasks or even causing damage in practical applications. However, the coefficients for eliminating joint drift are coupled with the equation constraints in Cartesian space under the existing scheme, which theoretically results in a paradox between zero joint drift and zero positional error in Cartesian space. To address the joint drift, a joint drift-free repetitive motion programme with position error feedback (JDF-RMPPEF) is presented and analyzed. The JDF-RMPPEF scheme decouples the joint error and position error, which enables the OMM to accurately perform the trajectory tracking and repetitive motion tasks. In addition, to suppress the disturbances and solve the JDF-RMPPEF problem accurately, an allowed nonconvex noise-resistant neural network (ANNRNN) model is proposed, which allows for a nonconvex activation function with noise suppression properties. Theoretical analysis demonstrates that the ANNRNN model exhibits global convergence and strong robustness in the presence of interference. Through examples and comparisons, the effectiveness and superiority of the JDF-RMPPEF scheme synthesized by the ANNRNN model are validated.

Various discrete-time zeroing neural network (DTZNN) models have been developed for solving dynamic constrained quadratic programming. However, two challenges persist within the DTZNN framework: first, the theoretical analysis of robustness in disturbance suppression remains insufficient; second, to the best of authors' knowledge, existing DTZNN models have yet to provide a theoretical proof of finite-step convergence. Inspired by the inherent robustness and finite-step convergence of discrete sliding-mode control based on the reaching-law, this article is the first work to integrate reaching-law theory into the DTZNN framework to address the aforementioned challenges, ensuring that the resulting DTZNN exhibits both robustness and finite-step convergence. In addition, a novel hyperbolic type reaching law (HTRL) is designed, which offers advantages in reducing the width of the quasi-sliding-mode region and suppressing chattering. The zeroing neural network (ZNN) based on this HTRL (HTRL-ZNN) is rigorously proven to exhibit effective disturbance suppression robustness and finite-step convergence, with an explicit expression provided for the convergence step length. Finally, the effectiveness and advantages of HTRL-ZNN in solving dynamic constrained quadratic programming are validated through both a numerical example and an application-oriented case.

This article proposes a novel data-driven iterative learning decoupling controller aimed at addressing the precise tracking issues in discrete-time nonlinear multi-input–multioutput (MIMO) repetitive systems, caused by difficult-to-characterize nonlinearity, multivariate coupling, and data noise pollution. First, a dynamic linearized data model (DLDM) is established, where the estimation of the pseudo-Jacobian matrix (PJM) in prior methods neglects noise suppression. To overcome this gap, a PJM adaptive robust estimation method using a noise-tolerant zeroing neural network (NN) is designed, guaranteeing residual-free convergence and enhancing robustness against data noise. Furthermore, an iterative sliding mode observer is developed to estimate the coupling between multiple variables, aiming to obtain a decoupled DLDM for synthesizing controllers. Next, the iterative learning controller (ILC) utilizing wavelet NNs is designed, and its convergence characteristics are analyzed theoretically. Integrating ILC with decoupling strategies simplifies controller design and provides new perspectives for analyzing the convergence of the MIMO system. Finally, the developed scheme’s performance are validated through case studies. Real-time feasibility is assessed using the dSPACE rapid prototyping system.

The coupling of multivariate repeated systems and the nonlinearity that is difficult to characterize through mechanisms, along with actuator constraints and data noise pollution, pose challenges in achieving precise tracking tasks. To address these issues, a novel data-driven robust iterative learning predictive control (ILPC) scheme is proposed. The contribution lies in its ability to achieve multivariable tracking without requiring any prior model information, all while effectively suppressing noise pollution and actively addressing actuator constraints. Specifically, a dynamic linearization data predictive model (DLDPM) is first obtained for system dynamic behavior prediction and controller synthesis. The estimation of the unknown pseudoJacobian matrix (PJM) in DLDPM was previously overlooked in terms of data noise suppression mechanisms. In this study, we utilize a noise-tolerant zeroing neural network (NT-ZNN) for its estimation. Theoretical analysis confirms that the PJM adaptive estimation law can achieve residue-free convergence and its robustness in noise suppression. Then, a constrained ILPC scheme is proposed, which transforms the multivariable tracking problem with actuator constraints into an iteration-varying quadratic programming problem with both inequality and equality constraints, which is solved using NT-ZNN. Theoretical proofs substantiate that a constrained ILPC scheme can achieve asymptotic convergence along the iterative axis. Finally, the proposed scheme is validated in a thermal management system for a proton exchange membrane fuel cell, showcasing the effectiveness in tracking tasks and handling actuator constraints in the presence of noise pollution.

Connectivity technologies in intelligent transportation systems offer unprecedented opportunities to enhance mobility, fuel economy, and safety for automotive systems. However, the uncertain driving behavior of surrounding vehicles in real-world traffic scenarios can significantly undermine these benefits. To tackle this challenge, this article develops a stochastic predictive-adaptive cruise control (P-ACC) system that effectively addresses uncertainties and automatically adapts to various driving scenarios. The proposed system employs a Gaussian process (GP)-based velocity predictor as its foundation, accurately capturing the driving dynamics of the preceding vehicle while accounting for prediction uncertainty using variances. The real-time feasibility is assessed in a dSPACE rapid prototyping system. In addition, the developed stochastic-model predictive control (S-MPC) approach incorporates the predicted velocity variance into the probabilistic chance constraints, conservatively narrowing the optimization space of the velocity planning domain, thereby enabling more reliable control. To further enhance the system’s performance in adapting to different driving conditions, a scenario-based parameter self-learning (PSL) technique is introduced in the S-MPC controller, utilizing Bayesian optimization (BO). Finally, the performance of the proposed controller is comprehensively evaluated by leveraging a high-fidelity simulator and on-board actual vehicle testing data. Simulation results demonstrate that the proposed method achieved a boost in tracking performance and driving comfort while maintaining fuel-saving benefits.

Traffic sign recognition is crucial for enhancing the safety and efficiency of intelligent transportation systems (ITS). This paper proposes a hybrid structure named DK-former, a lightweight and robust deep kernel Gaussian process transformer network, aiming to tackle non-linear small samples, high model parameters, and environmental variations, providing a distinctive solution that enhances robustness and adaptability in intelligent transportation systems. The whole structure is trained end-to-end based on the Bayesian framework. First, introducing convolutional random Fourier features facilitates effective non-linear sample feature mapping, capturing complex patterns inherent in traffic signs and enhancing computational efficiency. Second, the transformer model refines feature representations with its powerful self-attention mechanism to capture the global relationship between traffic sign pixels. Furthermore, the hybrid structure enhances the generalization capability of the model, allowing it to adapt to diverse scenarios, such as varying lighting conditions and weather fluctuations commonly encountered in real-world traffic environments. Extensive experiments have been conducted to validate the effectiveness of the proposed DK-former for traffic sign recognition. The model has achieved an outstanding recognition accuracy of 99.09% on the German Traffic Sign Recognition Benchmark (GTSRB) dataset, superior to current state-of-the-art deep kernel learning traffic sign models with only 8.3 million parameters. Experiments on Indian and Chinese datasets comprehensively demonstrate the generalization of the model across different geographies and environments. The code is publicly available at:https://github.com/w-tingting/DK-former.

Autonomous vehicle (AV) evaluation has been the subject of increased interest in recent years both in industry and in academia. This paper focuses on the development of a novel framework for generating adversarial driving behavior of background vehicle interfering against the AV to expose effective and rational risky events. Specifically, the adversarial behavior is learned by a reinforcement learning (RL) approach incorporated with the cumulative prospect theory (CPT) which allows representation of human risk cognition. Then, the extended version of deep deterministic policy gradient (DDPG) technique is proposed for training the adversarial policy while ensuring training stability as the CPT action-value function is leveraged. A comparative case study regarding the cut-in scenario is conducted on a high fidelity Hardware-in-the-Loop (HiL) platform and the results demonstrate the adversarial effectiveness to infer the weakness of the tested AV.

Under extreme conditions, vehicle states change rapidly between stable and unstable, resulting in dynamic requirements for the vehicle’s overall safety stability. Simultaneously, the coupled nonlinear characteristics of vehicle dynamics cannot be ignored in controller design. To address the above problems and improve vehicle longitudinal and lateral stability integrally, an envelope-based model predictive control (MPC) strategy with dynamic objectives is proposed for four-wheel independent motor-drive electric vehicles (4WIMD EVs). First, according to the current driving behavior and the collected road information, the envelope control regions concerning vehicle side-slip angle and yaw rate are obtained online, and divided into stable, critically stable, and instable regions with different safety requirements. Then, the safety dynamic requirements are constructed in the designed MPC-based control structure. A nonlinear vehicle dynamics model with a combined-slip tire model, which integrates the longitudinal and lateral dynamics, is utilized to predict vehicle states. The switching of requirements is reflected in the variation of weighting factors and constraint values. Finally, CarSim and Matlab/Simulink co-simulation, and hardware-in-the-loop simulation test results show better satisfactory performance in improving overall vehicle stability under extreme driving conditions.