For distributed-drive electric vehicles, torque vectoring control based on model predictive control (MPC) has emerged as a preferred strategy to achieve superior performance across diverse driving conditions. However, traditional methods heavily rely on high-order nonlinear models, which require extensive tuning to achieve optimal performance. This article proposes a gradient-based automatic hyperparameter tuning algorithm to enhance vehicle maneuverability and stability by optimizing the MPC objective function. First, a data-driven predictive motion control scheme is introduced, enabling accurate modeling of the real vehicle through a data-driven approach. Subsequently, a differentiable alternating direction method of multipliers (ADMM) is proposed to compute the gradients of state and control input trajectories with respect to MPC weights during optimization. Real vehicle experiments demonstrate that the method effectively captures the characteristics of various driving conditions and automatically adjusts parameters for adaptation. Compared to manual parameter tuning, the strategy improves maneuverability by 33.02%, lateral stability by 27.19%, and ride comfort by 20.78%. Furthermore, in comparison with existing gradient-based tuning methods, the differentiable ADMM achieves more stable online parameter tuning. These metrics are further enhanced by 19.62%, 5.23%, and 34.5%, respectively.

The pronounced nonlinearity of vehicle dynamics under extreme conditions presents a substantial challenge for motion control. The complexity of nonlinear modeling, parameter inaccuracies, and the high computational cost associated with high-order nonlinear control systems significantly hinder improvements in control stability and performance. To address these challenges, this article proposes an innovative data-driven model predictive control (DMPC) approach. A data-driven modeling (DM) technique is first introduced to construct a linear model based on step response input/output (I/O) data, with real-time compensation for model deviations using online data to more accurately capture the vehicle’s lateral and longitudinal dynamics. The stability of the model estimation error is rigorously analyzed, and a data-driven vehicle model with theoretical guarantees is established. A hierarchical stability control framework based on DM is then developed. The upper-level controller evaluates vehicle maneuverability and lateral stability to determine the required additional yaw moment, while the lower-level controller assesses longitudinal stability and traction, and incorporates lateral–longitudinal coupling to optimize motor torque distribution across the wheels. DM is validated using more than 25 000 s of low-adhesion road data, demonstrating significantly higher modeling accuracy than traditional nonlinear models (NLMs). Compared with the NLM-based controller, DMPC improves vehicle handling and lateral stability by 28.98% and 66.92%, respectively, while reducing driver workload by 33.72% in lateral tests. In longitudinal tests, the strategy increases vehicle acceleration by 58.65% and reduces yaw fluctuation by 79.5%. Finally, the robustness of DMPC is evaluated through parameter sensitivity analysis and variations in tire pressure.

Accurate joint estimation of vehicle lateral velocity and tire-road friction coefficient (TRFC) is critical for active safety systems. However, under near-limit conditions, strong nonlinear coupling and mutual error propagation severely degrade estimation accuracy. To address these challenges, this article proposes a novel uncertainty-aware hierarchical moving horizon estimation and extended Kalman filter (MHE–EKF) framework. This two-stage architecture employs a dual EKF (DEKF) to provide robust coarse priors, followed by a bidirectionally coupled MHE–EKF for high-precision joint estimation. The key novelty of this MHE–EKF structure lies in an embedded dual-innovation mechanism explicitly designed to manage uncertainty: first, an error-propagation model translates state errors into a high-confidence TRFC constraint interval, strictly bounding the MHE search space; second, a flag-driven adaptive arrival-cost strategy integrates multistage priors to accelerate mutual convergence. Vehicle experiments validate the algorithm. Compared to the conventional DEKF, the proposed framework reduces the lateral velocity mean absolute error to 28.8$\%$ and 42.1$\%$ of the DEKF baseline in dry-asphalt slalom and double-lane-change maneuvers, respectively, while concurrently decreasing the corresponding TRFC estimation errors from 5.73$\%$ to 0.11$\%$ and from 11.61$\%$ to 1.38$\%$. Extended evaluations on low-adhesion snow roads further confirm the algorithm's exceptional robustness and consistent accuracy across variable friction conditions.