Wind-aware deep reinforcement learning for autonomous UAV navigation without GPS/GNSS position feedback and visual perception
EXPERT SYSTEMS WITH APPLICATIONS, cilt.334, ss.1-31, 2026 (SCI-Expanded, Scopus)
- Yayın Türü: Makale / Tam Makale
- Cilt numarası: 334
- Basım Tarihi: 2026
- Doi Numarası: 10.1016/j.eswa.2026.134645
- Dergi Adı: EXPERT SYSTEMS WITH APPLICATIONS
- Derginin Tarandığı İndeksler: Applied Science & Technology Source, Academic Search Ultimate (EBSCO), Engineering Source (EBSCO), Scopus, Technology Collection (ProQuest), Aerospace Database, Science Citation Index Expanded (SCI-EXPANDED), Compendex, INSPEC, Public Affairs Index
- Sayfa Sayıları: ss.1-31
- Karadeniz Teknik Üniversitesi Adresli: Evet
Özet
The ability of unmanned aerial vehicles (UAVs) to maintain autonomous navigation under wind disturbances and in conditions where GPS/GNSS-based position feedback and visual perception are unavailable during flight requires the effective utilization of available environmental information. In this study, a Wind-Aware Deep Reinforcement Learning (WAGV-DRL) method is proposed to enable UAVs to sustain autonomous navigation under windy conditions in the absence of GPS/GNSS position feedback and visual perception during flight. In the proposed method, the initial and goal positions are predefined prior to flight; no visual or satellite-based positioning information is used during flight, while Wind Speed (WS) and Wind Direction (WD) data provided by an independent onboard sensing system are directly incorporated into the policies as state/observation inputs. The proposed method was evaluated under three main scenarios based on real flight data, including goal-oriented navigation under synthetic wind profiles and, using a real-world dataset comprising 1,639 observations (Dataset_1), the ability to maintain the reference flight trajectory and reach the goal position without deviation. In complementary experiments conducted under complete GPS/GNSS outage conditions, the mean position estimation error of 613.36 m obtained with the Basic Dead Reckoning (Basic DR) approach without wind correction was reduced by 99.89%, reaching 0.68 m with the wind-compensated approach (Wind-DR). DQN policies trained using Wind-DR-based position estimation achieved a 100% goal-reaching success rate (30/30) in independent training runs. Furthermore, in 30 closed-loop tests conducted on an independent dataset (Dataset_2) that had never been encountered during training, with the model weights kept fixed, a goal-reaching success rate of 86.67% was achieved without any fine-tuning. The findings clearly demonstrate the effectiveness of the proposed method and provide evidence that wind information can support autonomous navigation in the absence of GPS/ GNSS position feedback and visual perception by contributing to environmental awareness and position estimation.