A Hybrid Deep Learning Approach for Vehicle Wheel Slip Prediction in Off-Road Environments

Abstract

Wheel slip prediction is essential for safe navigation and optimal trajectory planning of ground vehicles, especially when traversing off-road on unpaved surfaces such as sand, gravel, or mud. However, calculating tire slippage precisely is cumbersome due to numerous sophisticated processes of measuring physical parameters related to the wheel-soil interaction. Most prior studies focused on developing slip prediction models suited for rovers or differential-drive robots, leaving car-like robots relatively overlooked. To this end, the present work develops a hybrid Deep Learning approach that addresses two key challenges: (i) identifying the terrain type on which the vehicle is driving, and (ii) estimating the wheel slip on uneven and unstructured surfaces. First, extensive data collection is carried out with an advanced simulator to construct a sufficiently descriptive dataset (504,000 samples) capturing various terrains, speed ranges, slopes, and maneuvers. Then, considering the close correlation between the terrain type and wheel slippage, we propose a lightweight convolutional neural network (CNN), referred to as TerrainNet, for accurate terrain classification. Lastly, leveraging the predictive power of TerrainNet, we train and compare the performance of several classical machine learning and deep learning regression techniques, namely multi-layer perceptron (MLP), random forest (RF), and extreme gradient boosting (XGB). The simulation results indicate that the proposed CNN can accurately discriminate the terrain (mean accuracy > 99%), enabling precise wheel slip estimations with the employed machine learning models (average root mean square error < 0.03).