Abstract
Canopy occlusion, dynamic vegetation, structural degeneracy, and implicit terrain risks make stable localization and task-adaptive planning difficult for resource-constrained orchard robots. This paper proposes CitraNav, a lightweight navigation method for global navigation satellite system (GNSS)-denied orchards. It separates stable geometric evidence for localization from short-lived semantic evidence for planning, preventing semantic observations from accumulating in the global map. For localization, a hierarchical voxel map selects its resolution according to local structure and light detection and ranging (LiDAR) sampling characteristics, while cross-frame reliability and observability constraints suppress updates from transient vegetation and weakly observable directions. For planning, synchronized color and depth observations form a local semantic risk point cloud. A model predictive path integral (MPPI) planner combines task-dependent semantic costs with exact-footprint collision checking against currently detected obstacles. In simulation, CitraNav achieved a mean translational localization root mean square error (RMSE) of 0.075 m. Compared with geometric point-cloud planning, semantic planning reduced the collision rate by 71.4% and increased weed coverage 4.72-fold. Across 14 real-world sequences spanning farm-road, lawn, forest, and orchard environments, CitraNav achieved mean translational and heading RMSEs of 0.151 m and 1.13°, respectively, while using 72.3–87.4% fewer geometric map cells than the comparison methods. The complete perception–planning pipeline operated at 20.3–32.7 frames per second on an edge platform. These results suggest that CitraNav offers a balanced approach to localization stability, task-adaptive planning, and computational efficiency in complex orchard navigation.