SRPG: Semantic Risk Preference-Guided Reinforcement Learning for UAV Navigation
IEEE Robotics and Automation Letters · 2026-12 · 11 (12) · 14276-14283
DOI: 10.1109/lra.2026.3746907 ↗Kurz gefasst
Fachartikel in der Zeitschrift IEEE Robotics and Automation Letters, 2026-12. Verfasst von Mengxian Li und 5 weiteren Autor:innen. Die Arbeit nennt 29 Literaturangaben. Gefördert von: Strategic Priority Research Program of CAS, Beijing Natural Science Foundation.
Abstract (Originalsprache)
Autonomous uncrewed aerial vehicles are increasingly deployed in cluttered environments, where safe local navigation depends not only on geometric clearance to surrounding objects but also on the semantic risk associated with those objects. Existing semantic risk methods commonly encode category-dependent risk as costmaps, risk maps, or other spatial cost fields, and then translate them into motion through a separate planning or optimization module. Meanwhile, learning-based UAV navigation policies mainly emphasize geometry-based collision avoidance, making it difficult to express category-dependent preferences when multiple collision-free directions have comparable clearance. To address this problem, we propose SRPG, a semantic risk preference-guided reinforcement learning framework for local UAV navigation. SRPG converts semantic observations into a compact LiDAR-aligned semantic risk observation with barrier-aware angular diffusion, allowing category-dependent risk to be used directly by the policy without maintaining a persistent semantic map or a grid-based local costmap for planning. The diffused semantic risk observation is used as policy input, and a semantic-action-risk reward directly evaluates the semantic risk associated with the motion direction selected by the current action, encouraging low-risk local motion choices under comparable geometric conditions. Simulation experiments show that SRPG improves the success rate over the strongest learning-based baseline by 8.0 and 9.6 percentage points in the static and dynamic basic navigation environments, respectively. In semantic risk preference tests, SRPG consistently selects the lower-risk side in different scenarios, indicating that the learned policy develops a consistent semantic risk preference. Real-world indoor flight experiments on a physical quadrotor further demonstrate that SRPG can be deployed without real-world fine-tuning and prefers the lower-risk side when navigating around risk-labeled objects.
Förderung
- Strategic Priority Research Program of CAS
- Beijing Natural Science Foundation
Zitieren
APA
Li, M., Wang, Q., Wu, Y., He, N., Li, T., & Xu, Y. (2026). SRPG: Semantic Risk Preference-Guided Reinforcement Learning for UAV Navigation. IEEE Robotics and Automation Letters, 11(12), 14276-14283. https://doi.org/10.1109/lra.2026.3746907
Vancouver
Li M, Wang Q, Wu Y, He N, Li T, Xu Y. SRPG: Semantic Risk Preference-Guided Reinforcement Learning for UAV Navigation. IEEE Robotics and Automation Letters. 2026;11(12):14276-14283. doi:10.1109/lra.2026.3746907
BibTeX
@article{Li2026,
author = {Li, Mengxian and Wang, Qi and Wu, Yiqun and He, Nan and Li, Tingting and Xu, Yongjun},
title = {SRPG: Semantic Risk Preference-Guided Reinforcement Learning for UAV Navigation},
journal = {IEEE Robotics and Automation Letters},
year = {2026},
volume = {11},
number = {12},
pages = {14276--14283},
doi = {10.1109/lra.2026.3746907}
}
EUVIMED