W-RIZZ：移动机器人中相对可遍历性估计的弱监督框架

在非结构化环境中成功部署移动机器人 requires understanding environment and terrain to avoid hazards, getting stuck, and collisions. Traversability estimation—predicting where a robot can travel—is a key approach. Existing geometric methods ignore semantic considerations, while semantic segmentation approaches involve tedious labeling. Recent self-supervised methods reduce labeling burden but require additional data/models and struggle to explicitly label untraversable areas. To address these limitations, we introduce a weakly-supervised method for relative traversability estimation. Our method involves manually annotating relative traversability of a small number of point pairs, significantly reducing labeling effort compared to traditional segmentation-based methods and avoiding self-supervised methods' limitations. We further improve performance through a novel cross-image labeling strategy and loss function. We demonstrate viability through deployment on a mobile robot in outdoor environments.