Real-time Terrain Analysis for Off-road Autonomous Vehicles

One challenge for autonomous vehicle control is the variation in road roughness which can lead to deviations from the intended course of loss of road contact while steering. The aim of this work is to develop a real-time road roughness estimation system using a Bayesian-based calibration routine that takes in axle accelerations from the vehicle and predicts the current road roughness of the terrain. The Bayesian-based calibration method has the advantage of providing posterior distributions and thus giving a quantifiable estimate of the confidence in the prediction that can be used to adjust the control algorithm based on desired risk posture. Within the calibration routine, a Gaussian process model is first used as a surrogate for a simulated half-vehicle model which takes vehicle velocity and road surface roughness (Gd) to output the axle acceleration. Then the calibration step takes in the observed axle acceleration and vehicle velocity and calibrates the Gaussian process model to best fit this data. The final result is the posterior distribution of the road surface roughness. To train the Gaussian process model, the half-vehicle model is used to collect vertical axle acceleration data over a range of velocities and road surface roughness levels using Latin Hypercube sampling. The Bayesian-based calibration method was then implemented in the loop with a simplex controller to update the velocity limits based on the predicted road surface roughness. To demonstrate the effectiveness of the control algorithm, a stochastically generated surface with different regions of varying road roughness was utilized to test the algorithm's ability to characterize road roughness in real-time, thus allowing the display of the simplex control strategy that enhances safety in autonomous vehicle operation. The proposed algorithm has the potential to mitigate the risks associated with road surface roughness, ensuring a safer and more efficient operation for autonomous vehicles.