The rapid expansion of unmanned aerial vehicles (UAVs) across diverse application domains has underscored the need for reliable autonomy in complex and dynamic environments. To advance toward this goal, in recent years, learning-based control strategies have emerged as a promising alternative, offering adaptability and decision-making capabilities beyond those of conventional model-based ones. Bearing this in mind, the proposed article examines reinforcement learning methodologies for controlling UAVs, with particular emphasis on commonly used virtual environments, benchmark tasks, and the challenges of bridging the gap between simulation and real-world deployment. Therefore, key limitations, including high computational demands, reliance on extensive training data, and reduced robustness under environmental variability, are critically analyzed from a practical implementation perspective. Rather than adopting an algorithm-centric viewpoint, this work aggregates existing knowledge and categorizes learning-based approaches by their level of control abstraction and their treatment of safety and stability, thereby identifying the key factors limiting large-scale real-world deployment and the trends shaping intelligent controllers.
Memlikai et al. (Tue,) studied this question.