Estimating the Road Friction Coefficient Dynamically using Machine Learning

In autonomous vehicles, it is vital for the vehicle to drive in a manner that ensures the driver is comfortable and does not feel compelled to intervene. Hence, the system must consider important environmental factors, such as the road friction coefficient, which quantifies the grip experienced by the tire on a road. An inaccurate estimation of road friction can lead to safety risks or discomfort for the driver, as the system would be overestimating or underestimating the tire’s grip on the road and the vehicle’s response to control inputs. Therefore, in the context of Advanced Driver Assistance Systems (ADAS), dynamically estimating the road friction coefficient can significantly improve the safety and comfort of autonomous driving functions. However, estimating the friction coefficient dynamically requires complex mathematical modelling of nonlinear relationships that are challenging to solve numerically or analytically. This, coupled with the need for real-time estimation presents a noteworthy challenge in practical systems. Even the State-of-the-art driving functions often assume a fixed friction value or operate within a friction range. On the other hand, machine learning-based approaches can model complex relationships using large amounts of data and estimate the road friction coefficient in real-time. In this paper, we propose a multivariate time-series model to dynamically estimate the friction coefficient. The model is trained on a synthetic dataset comprising vehicle dynamics data for cars driven on roads with different friction levels. We evaluate multiple model architectures and hyperparameter configurations against metrics like accuracy and inference time to identify the best-performing model. Furthermore, the models are also evaluated with real-world data, where input data is only available at a lower frequency, to validate their performance in scenarios akin to real driving contexts.