Self-supervised Traversability Estimation in Off-road Environments Using Geometric Superpixels and Semantic Features

This study presents a self-supervised traversability estimation algorithm and an integrated navigation system for unmanned ground vehicles (UGVs) operating in unstructured off-road environments without the use of pre-built maps. Conventional approaches, such as model-predictive-control based traversability modeling and LiDAR-based obstacle detection, often suffer from high computational cost, sensitivity to parameter tuning, and inaccurate terrain classification. RGB-based semantic segmentation alone is also vulnerable to illumination changes and color ambiguities, leading to unreliable performance in unstructured terrains. To address these challenges, the proposed method fuses RGB-d sensory data, semantic features, and real-world driving experience within a self-supervised learning framework. Depth SLIC- based superpixel segmentation reduces dependence on RGB appearance and provides geometrically consistent regions, while superpixel-level aggregation of semantic probabilities creates compact and discriminative feature vectors. Traversability labels are automatically generated from driving data by analyzing discrepancies between commanded and actual velocities, and roll and pitch variations, and are projected onto the image plane for dense supervision without manual annotation. An multilayer-perceptron based model predicts traversability while maintaining consistency among semantic, geometric, and motion features. The predicted traversability is transformed into a bird's-eye-view local cost map and integrated with the ROS2 NAV2 DWB controller, enabling real-time navigation without using a global map. Experiments on a custom off-road UGV platform demonstrate reliable obstacle avoidance and traversable region estimation, highlighting the method's effectiveness in enhancing autonomy and robustness in extreme off-road conditions.