Toward GPS-independent ground vehicle control: visual navigation and traversability-aware planning in unstructured environments

This paper presents a novel global positioning system (GPS)-independent visual navigation and motion control framework for unmanned ground vehicles operating in unstructured environments. While autonomous navigation has achieved significant milestones in structured environments by leveraging the synergy of GPS and high-definition maps, unstructured terrains—including forests, agricultural expanses, and disaster zones—remain a formidable challenge due to GPS signal degradation or total unavailability, coupled with complex and cluttered environmental conditions. To address this issue, we propose leveraging onboard monocular RGB camera for visual navigation, thus eliminating the need for GPS or global maps. Our framework utilizes NoMaD, a goal-conditioned diffusion model that generates multiple candidate paths based solely on visual input, enabling trajectory sampling and planning without GPS. To further improve performance in obstacle-dense environments, we introduce a traversability-aware risk model to assess the risk for specific waypoint on each sampled trajectory, selecting the optimal target point for pure pursuit tracking to ensure safe navigation. We validate the proposed approach through real-vehicle experiments in unstructured environments, demonstrating its effectiveness in achieving reliable navigation, efficient obstacle avoidance, and high traversal performance, all without any dependency on GPS. Demonstration video can be found in the supplementary material.