Physics-informed and machine learning approaches for real-time road surface grip classification

A reliable road grip estimate is of paramount importance to improve vehicle safety, control performance, and adapt driving strategies, especially in rapidly changing environmental conditions. However, real-time grip friction estimation remains challenging due to limited excitation of tyre–road forces, model uncertainties, sensor noise, and rapidly varying road conditions. To overcome this issue, this work presents and experimentally evaluates two complementary strategies for real-time road surface grip classification based on signals typically available in standard vehicle instrumentation. First, a model-based strategy is presented that relies on vehicle dynamics to estimate longitudinal tyre–road interaction forces, while a second supervised learning-based method is directly applied to measured signals using physics-informed features. To address the limited availability of force measurements in production standard vehicles, variants of both approaches are considered in which the classification is performed without assuming the availability of feedback forces, which are typically not measured. The proposed methods have been validated on experimental data collected on real test tracks under multiple road surface conditions, including dry and wet asphalt, snow, and ice. Results demonstrate the effectiveness of both approaches. The physics-based one provides accurate system description and improved interpretability, although its performance degrades in the force reconstruction phase. Conversely, the learning-based method achieves high classification accuracy and robustness to noise and modeling uncertainties, reaching an F1-score of 98.2% with decision latency below 50 ms. These results demonstrate the feasibility of reliable grip classification using minimal sensing, supporting practical deployment in ADAS applications.