Learning-based vision-guided precision landing for autonomous UAVs: comparative evaluation of reinforcement learning and evolutionary controller optimization
This paper presents a vision-guided framework for the terminal landing phase of an unmanned aerial vehicle (UAV) and compares two approaches to adaptive control: model-free policy learning with tabular Q-learning and genetic-algorithm optimization of a structured feedback controller (GA-PID), which in the investigated implementation operates in proportional–derivative form, which in the investigated implementation operates in proportional–derivative form. Both controllers share the same perception pipeline, in which a low-cost downward-facing camera detects an ArUco fiducial marker, and require neither GNSS nor external positioning during this phase. The controllers were trained and optimized in a ROS 2/Gazebo simulation and transferred to a physical multirotor UAV, so that both are evaluated with the same sensing, platform, landing task, and evaluation protocol. The results show that tabular Q-learning acquires a functional landing policy despite its compact state representation. In real-flight trials, the GA-PID controller landed significantly faster, whereas the differences in final accuracy and success rate were not statistically significant with ten trials per controller. In simulation, the GA-PID controller produced smoother trajectories, and its practical insensitivity to a change in illumination was confirmed by equivalence testing. The findings characterize the two specific implementations under the investigated conditions rather than ranking reinforcement learning and evolutionary optimization in general.The results show that tabular Q-learning acquires a functional landing policy despite its compact state representation. In real-flight trials, the GA-PID controller landed significantly faster, whereas the differences in final accuracy and success rate were not statistically significant with ten trials per controller. In simulation, the GA-PID controller produced smoother trajectories, and its practical insensitivity to a change in illumination was confirmed by equivalence testing. The findings characterize the two specific implementations under the investigated conditions rather than ranking reinforcement learning and evolutionary optimization in general. The framework provides interpretable baselines and a reproducible basis for future research on vision-guided UAV landing.
Authors
- Leszek Ambroziak (ORCID: https://orcid.org/0000-0001-7612-768X)
- Arkadiusz Bożko (ORCID: https://orcid.org/0000-0002-3336-0315)
- Mirosław Kondratiuk (ORCID: https://orcid.org/0000-0002-5061-6781)
- Suryansh Sharma (ORCID: https://orcid.org/0000-0001-7534-6848)
- Łukasz Łobacz
- R. R. Venkatesha Prasad
Institutions
- Bialystok University of Technology (PL)
- Delft University of Technology (NL)
Publication Details
- Journal
- Frontiers in Neurorobotics
- Published
- 2026-11-19
- DOI
- https://doi.org/10.3389/fnbot.2026.1974605
- Primary Topic
- Aerospace Engineering and Control Systems
- Type
- article
- Field-Weighted Citation Impact
- 0.00