Self-supervised Traversability Learning With Image Segmentation Foundation Model for Mobile Robots

Mobile robots' autonomous navigation in diverse environments critically depends on accurate traversability estimation, which traditionally relies on predefined parameters. This approach, while controllable, lacks flexibility in complex, dynamic real-world settings. To address this limitation, we propose a novel method integrating self-supervised traversability learning with an image segmentation foundation model. Our approach leverages the robot's actual traversal data as baseline annotations for the foundation model, enabling traversability learning for similar, unexperienced terrains. This method enhances the robot's ability to generalize traversability estimation beyond directly experienced environments. By utilizing the image segmentation foundation model, our system can extract traversable areas from visual data, correlating them with the robot's mobility capabilities. Experimental results demonstrate the effectiveness of our approach in data extraction and learning on similar, untraversed surfaces, thus expanding the robot's operational range and adaptability. Our framework offers a more robust and adaptable solution for mobile robot navigation, particularly in terrains where traversability assessment is crucial. This research contributes to the advancement of autonomous navigation systems, potentially improving the efficiency and safety of mobile robots in diverse and challenging environments.