|
|
|
| A UAV Visual Reinforcement Learning Navigation Framework for Continuous Control |
| CHEN Lijun |
| School of Software and Artificial Intelligence， Software Engineering Institute of Guangzhou，
Guangzhou 510990， Guangdong， China |
|
|
|
|
Abstract When performing visual navigation tasks in complex unknown environments, unmanned aerial vehicles (UAVs) face various security threats such as sensor noise, GPS spoofing, and dynamic obstacles, and existing methods struggle to simultaneously ensure robustness and real-time performance. This paper proposes SecNav, a multi-level visual navigation framework for safety enhancement. The framework integrates adversarial input detection, multimodal feature extraction, and safe reinforcement learning decision-making to achieve end-to-end secure navigation. Firstly, an anomaly detection module based on a variational autoencoder (VAE) is constructed, which utilizes reconstruction error and Kullback-Leibler (KL) divergence to identify spoofing attacks and anomalous data. Secondly, a multimodal feature extraction network integrating an attention mechanism and inertial measurement unit (IMU) information is designed to enhance perceptual robustness under visual degradation conditions. Finally, a constrained Markov decision process (CMDP) model with safety constraints is introduced, and the Lagrangian method is employed to optimize the balance between reward and safety constraints, achieving safety-aware decision and control. Experiments on the AirSim simulation platform and a self-built adversarial dataset demonstrate that SecNav reduces trajectory deviation by 72.5% under GPS spoofing attacks, improves the success rate of dynamic obstacle avoidance by 29.8%, achieves an attack detection accuracy exceeding 92%, and simultaneously satisfies the real-time processing requirement of 30 frames per second, validating the significant advantages of the proposed method in terms of safety and real-time performance.
|
|
Received: 17 September 2025
Published: 04 August 2026
|
|
|
|
|
|
| [1] |
WANG Jianyuan, CHEN Xiaotong, ZHANG Yue, SUN Junge, SHI Donghao, CHEN Jinbao. Urban Target Detection Algorithm Based on Multi-Modal Fusion of UAV[J]. Air & Space Defense, 2024, 7(1): 32-39. |
| [2] |
ZHAO Huipan, LIU Huanyu. Low Signature Target Group Detection Based on Multimodal Data Learning Fusion Network[J]. Air & Space Defense, 2021, 4(3): 41-47. |
|
|
|
|
2026, Vol. 9 
