Deep learning-based robust autonomous navigation of aerial robots in dense forests

Autonomous aerial navigation in dense natural environments remains challenging due to limited visibility, thin and irregular obstacles, GNSS-denied operation, and frequent perceptual degradation. This work presents an improved deep learning–based navigation framework that integrates semantically enhanced depth encoding with neural motion-primitive evaluation for robust flight in cluttered forests. Several modules were incorporated on top of the original sevae-ORACLE algorithm, to address limitations observed during real-world deployment, including lateral control for sharper maneuvering, a temporal-consistency mechanism to suppress oscillatory planning decisions, a stereo-based VIO solution for drift-resilient state estimation, and a supervisory safety layer that filters unsafe actions in real time. A depth-refinement stage was additionally included to improve the representation of thin branches and reduce stereo noise, while GPU optimization increased onboard inference throughput from 4 Hz to 10 Hz. To contextualize system performance, a comparative evaluation was conducted against several existing learning-based navigation methods tested under identical environmental conditions and hardware constraints. The proposed approach demonstrated consistently higher success rates, more stable trajectories, and improved collision avoidance, particularly in highly cluttered forest settings. The system was deployed on a custom quadrotor in three boreal forest environments. In moderate-clutter and dense environments, it achieved fully autonomous completion in all 15/15 flights, while in a highly dense underbrush setting it succeeded in 12/15 trials, representing a substantial improvement over the previous baseline, which completed only 5/15 comparable missions. These findings demonstrate that the proposed system provides a measurable improvement in reliability and safety over existing navigation methods in complex natural environments.