Physics-guided online adaptive vehicle dynamics modeling via decoupled residual learning
Accurate prediction of vehicle dynamics is critical for reliable motion planning and control of autonomous vehicles. However, physics-based models often struggle to represent the strong nonlinearity and coupling inherent in vehicle dynamics, while purely data-driven models frequently deteriorate when operating conditions deviate from the training domain. Physics-informed neural networks (PINNs) alleviate this issue by incorporating physical priors, but their performance can still be constrained by incomplete physical equations. To address these limitations, this paper proposes a Decoupled Hybrid Residual Model (DHRM) for online adaptive prediction of vehicle dynamics. The overall modeling residual is decomposed into a static structural component and a dynamic stochastic component, which are compensated through an offline channel and an online channel, respectively. The offline static channel combines Kolmogorov-Arnold networks (KAN) and long short-term memory (LSTM) networks to learn the static residual arising from structural simplification and parameter uncertainty. Concurrently, the online dynamic channel employs a sparse Gaussian process (SGP) to adaptively compensate the dynamic residual induced by changing operating conditions and external disturbances. Extensive simulation and scaled vehicle experiments validate the proposed framework. Under an abrupt friction change, DHRM-SGP reduces the root mean square error (RMSE) of lateral velocity and yaw rate by 67.5% and 70.5%, respectively, compared with the nominal physical model. The deployed model achieves an average inference latency of 4 . 6036 ms per sample on the Jetson Orin NX, supporting the feasibility of deploying the prediction module on onboard hardware. These results demonstrate that the proposed architecture effectively improves prediction accuracy, robustness, and implementation feasibility under varying driving conditions.