Terrain-Inspired Bumpiness Prediction for Off-Road Autonomous Driving

With the increasing demands for stability and comfort of unmanned ground vehicles (UGVs) driving in off-road environments, there is a higher requirement for bumpiness prediction in various terrains. Bumpiness prediction can accurately characterize the vehicle’s dynamic performance when traversing terrains, but it remains vulnerable to exteroceptive sensing errors caused by ego-motion uncertainty. To address these requirements, we propose a self-supervised learning framework for off-road bumpiness prediction. The framework is designed to achieve stable terrain perception under inevitable vehicle vibrations. Bayesian generalized kernel (BGK) inference is applied to compensate for perceptual blind spots caused by rapid maneuvers of the UGV. For bumpiness prediction, we introduce a convolutional Long Short-Term Memory (conv-LSTM) network that takes the vehicle speed and feature images extracted from light detection and ranging (LiDAR) point clouds as inputs, with spatially corresponding Inertial Measurement Unit (IMU)-derived bumpiness values serving as self-supervised labels. Real-world experimental results demonstrate that the proposed framework can accurately predict bumpiness values that consistent with the vehicle’s dynamic performance, even under ego-motion uncertainty.