Experimental Evaluation of Gaussian Process-Based Residual Models for Vehicle Dynamics Compensation

Gaussian process (GP)-based residual models have demonstrated remarkable effectiveness and practical value in compensating for vehicle dynamics model errors. The selection of input features for GP is critical in determining the accuracy of model error prediction. This paper investigates the influence of different input feature selection on the performance of residual models. To this end, three GP-based residual models employing different input features are developed to enhance the accuracy of a nominal vehicle dynamics model. The resulting hybrid models are subsequently integrated into a model predictive controller. The performance of the proposed models and controllers is rigorously validated through extensive openloop and closed-loop experiments within a high-fidelity simulation environment. Experimental findings highlight the dependence of input feature selection on vehicle operating conditions and demonstrate that appropriately simplifying input features can effectively reduce model complexity while preserving predictive accuracy and control performance.