Road Grip Uncertainty Estimation Through Surface State Segmentation

TL;DR

This work addresses uncertain grip estimation under slippery road conditions for autonomous driving. It benchmarks standard uncertainty regression methods and introduces Grip via Road State (GvRS), a novel approach that leverages pixel-wise road-surface state segmentation to produce a grip probability distribution by combining surface-state probabilities with class-specific grip distributions. Experiments on a Finland-driving dataset show that while traditional methods can efficiently predict mean grip, GvRS offers superior robustness in uncertainty estimation, particularly in out-of-distribution scenarios, albeit sometimes at the cost of wider intervals. The study highlights practical implications for safer vehicle control and points toward hybrid methods that merge direct grip regression with surface-state-informed uncertainty modeling.

Abstract

Slippery road conditions pose significant challenges for autonomous driving. Beyond predicting road grip, it is crucial to estimate its uncertainty reliably to ensure safe vehicle control. In this work, we benchmark several uncertainty prediction methods to assess their effectiveness for grip uncertainty estimation. Additionally, we propose a novel approach that leverages road surface state segmentation to predict grip uncertainty. Our method estimates a pixel-wise grip probability distribution based on inferred road surface conditions. Experimental results indicate that the proposed approach enhances the robustness of grip uncertainty prediction.

Figures (6)

• Figure 1: Two ways of approaching the grip uncertainty estimation problem: (A) using a regression model with predictive uncertainty or (B) leveraging the proposed method, which combines road surface state segmentation with grip probability distributions for each surface type to estimate grip and its uncertainty.
• Figure 2: Stacked histogram of grip values for different road surface state classes in the training set.
• Figure 3: Normalized grip histograms (dashed line) and corresponding fitted grip distributions (solid line) for each road surface state class. Vertical lines show the knots of the piecewise linear functions approximating the grip distributions.
• Figure 4: Scatter plots of the test set samples showing the relation between ground truth grip and the corresponding predicted 5th percentile for each model. The average percentage of points over / under the 5th percentile limit are shown in each title. Each point represents the averaged metrics in one test set sample, shown in green class if 90% of the ground truth grip measurements in the sample are over the 5th percentile limit, and shown in red class otherwise.
• Figure 5: Visualizations of grip and grip uncertainty output on the test set images. The first image for each example shows the ground truth data from the road weather sensor. For each model output, the upper image shows the predicted grip distribution mean and the lower image shows the distance between the predicted 5th percentile limit and the predicted grip distribution mean. The road area is manually segmented in the images for clarity.