FW-MonoAvoid-Net: Real-Time Monocular 3D Obstacle Avoidance for Fixed-Wing UAVs in Unmapped Environments

Autonomous flight of UAVs in unmapped environments remains a hot topic in research. While vision- and LiDAR-based autonomy has advanced rapidly for multirotor platforms, research on low-altitude autonomous flight for fixed-wing UAVs is still limited. To address obstacle avoidance in unmapped and complex 3D environments for fixed-wing UAVs, this paper proposes a neural network named FW-MonoAvoid-Net. This network is an end-to-end monocular-vision-based framework for fixed-wing UAVs to accomplish environment awareness and obstacle avoidance. Firstly, we develop a simulation framework to collect safe flight-trajectory data across diverse flight conditions, which is employed to train the proposed FW-MonoAvoid-Net. The proposed flight-learning framework adopts a lightweight visual perception network and a kinematics-aware state fusion module to predict the trajectory. At the output stage, an explicit physics-constrained decoder is embedded to strictly confine predicted trajectories within the feasible flight envelope of fixed-wing UAVs. Experiments are conducted in aerial (simulated spherical obstacles), canyon, urban (simulated high buildings), and New York Manhattan city scenarios. In the Manhattan city test, it achieves an 82.26% success rate at approximately 105 m flight altitude and 20 m/s cruise speed. Manhattan city is not included in the training datasets, FW-MonoAvoid-Net did not “see” similar scenarios in the training process. The testing results demonstrate its strong generalization and zero-shot deployment capability across unmapped environments within the simulation setting. FW-MonoAvoid-Net is a “smaller” model which has fewer than 3 million parameters. A Jetson Orin NX 16GB platform was employed to test the computation efficiency, and the trajectory-generation module reaches an average inference latency of 5.21 ms (about 190 Hz). These results show that, under limited computation and power consumption, the proposed method maintains both high success rates and excellent real-time performance. FW-MonoAvoid-Net is expected to help develop a fully autonomous fixed-wing flight in unmapped 3D spaces.