Autonomous Driving on Forest Roads under Unreliable GNSS

Reliable autonomous navigation in forest environments remains challenging due to their semi-structured road geometry, vegetation-induced occlusions, and unreliable GNSS-based localization. Traditional waypoint-based navigation methods depend on accurate global positioning and often fail in forests, where dense vegetation causes signal attenuation and multipath reflections, leading to large localization errors. This paper proposes a hierarchical navigation framework that integrates localization, perception, and planning in a novel way to enable autonomous driving on forest roads under unreliable GNSS conditions. The system employs LiDAR-IMU odometry for local motion estimation, a global localization module that monitors GNSS consistency through trajectory validation, and a high-level road network representation obtained from satellite imagery on which a global route planner computes the desired path. A stereo-vision-based perception module detects visible roads and reconstructs their 3D geometry, while a road selector compares these candidates with the planned path to identify the correct branch at intersections. Finally, a Model Predictive Controller (MPC) tracks the selected road using odometry feedback while enforcing vehicle dynamics constraints. Experimental results on real forest roads demonstrate that the proposed framework maintains consistent localization and stable control during extended GNSS degradation, achieving robust road following and reliable intersection handling.