Research map / Vision-language and multimodal / Robot action and navigation
Angle Robustness Unmanned Aerial Vehicle Navigation in GNSS-Denied Scenarios
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes an end-to-end angle-robust vision-based UAV point-to-point navigation model that directly predicts direction angles from recent frames, plus the UAV_AR368 dataset and a Google Earth-based simulation testing tool (SFTI).
Problem
Vision-based UAV point-to-point navigation in GNSS-denied settings suffers from flight deviation and cumulative error because existing classification- and matching-based methods first estimate position and then compute a direction angle.
Approach
Directly regress the direction angle (as sin/cos) end-to-end from a sequence of historical frames and their positions, using a MobileNetV3 feature extractor, a decoder-only Transformer, a HardTanh-based robust activation head, and auxiliary position-classification heads.
Key results
- The proposed model reaches SR@25 of 100.0% in ideal and 67.5% in disturbed environments, reported as 26.0% and 45.6% above baselines.
- Removing the Current or Next Position Prediction Head drops SR@25 in disturbed settings to 32.5% and 47.5%; removing the Robust Activation Module drops it to 21.9%.
- Classification-based methods achieve 0.0% success rate; matching-based baselines reach around 3%.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| directly predict direction angles with the proposed angle-robust method | UAV point-to-point navigation · UAV_AR368 ideal | SR@25: 26.0% higher than that of the baselines | Helped |
| directly predict direction angles with the proposed angle-robust method | UAV point-to-point navigation · UAV_AR368 disturbed | SR@25: 45.6% higher than that of the baselines | Helped |
| add the Displacement Guidance Module (ablation) | Angle robustness model · UAV_AR368 ideal and disturbed | SR@25: 6.0% -> 100.0%; 3.1% -> 67.5% | Helped |
| add the Robust Activation Module (ablation) | Angle robustness model · UAV_AR368 ideal and disturbed | SR@25: 74.0% -> 100.0%; 21.9% -> 67.5% | Helped |
| add the Adaptive Feature Enhance Module (ablation) | Angle robustness model · UAV_AR368 ideal and disturbed | SR@25: 0.0% -> 100.0%; 0.0% -> 67.5% | Helped |
| add the Cross-knowledge Attention-guided Module (ablation) | Angle robustness model · UAV_AR368 ideal and disturbed | SR@25: 0.0% -> 100.0%; 0.0% -> 67.5% | Helped |
| add the Current Position Prediction Head (ablation) | Angle robustness model · UAV_AR368 ideal and disturbed | SR@25: 63.0% -> 100.0%; 32.5% -> 67.5% | Helped |
| add the Next Position Prediction Head (ablation) | Angle robustness model · UAV_AR368 ideal and disturbed | SR@25: 68.0% -> 100.0%; 47.5% -> 67.5% | Helped |
| remove both Current and Next Position Prediction Heads (ablation) | Angle robustness model · UAV_AR368 ideal and disturbed | SR@25: 100.0% -> 0.0%; 67.5% -> 0.6% | Hurt |
| vary the number of candidate images for FSRA (25, 33, 41) (ablation) | FSRA · UAV_AR368 ideal and disturbed | SR@25: 3.0%, 3.0%, 2.0%; 0.6%, 1.9%, 0.6% | Mixed |
| vary the number of candidate images for RK-Net (17, 25, 33) (ablation) | RK-Net · UAV_AR368 ideal and disturbed | SR@25: 0.0%, 1.0%, 0.0%; 0.0%, 0.6%, 0.0% | Mixed |
| vary the number of candidate images for LPN (9, 17, 25) (ablation) | LPN · UAV_AR368 ideal and disturbed | SR@25: 1.0%, 3.0%, 1.0%; 1.3%, 0.0%, 0.6% | Mixed |
Surprising findings
- Classification-based methods fail all flights with 0.0% success rate, as they misclassify initial images.
Limitations
- Model does not reach 100% success in disturbed environments; authors attribute this to limited training samples, limited route types, and possible over/under-fitting.
- Evaluated on a limited number of routes and only in a simulated Google Earth setting rather than real flights.
Experimental setup
Datasets: UAV_AR368 ImageNet University-1652 CVUSA
Benchmarks: UAV_AR368
Metrics: SR@25 SR@50 MEPE MRE MRD inference time storage
Baselines: MobileNetV3 ResNet18 ViT-B/16 FSRA RK-Net LPN Backbone-FC
Builds on: MobileNetV3 Transformer Task-Uncertainty Loss imgaug
Closest papers by meaning
- Game4Loc: A UAV Geo-Localization Benchmark from Game Data (AAAI 2025 · similarity 0.84)
- Weather-Robust Cross-View Geo-Localization via Prototype-Based Semantic Part Discovery (NeurIPS 2026 · similarity 0.83)
- Towards Realistic UAV Vision-Language Navigation: Platform, Benchmark, and Methodology (ICLR 2025 · similarity 0.82)
- Is your VLM Sky-Ready? A Comprehensive Spatial Intelligence Benchmark for UAV Navigation (CVPR 2026 · similarity 0.82)
- SOAR: Regression-based LiDAR Relocalization for UAVs (NeurIPS 2026 · similarity 0.81)
- SatNav: A Scalable Benchmark for Long-Horizon UAV Vision-Language Navigation from Satellite Imagery (NeurIPS 2026 · similarity 0.81)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.