Drones are increasingly employed in critical applications, yet ensuring their safe operation in dynamic and unpredictable environments remains a challenge. This thesis examines the use of reach-avoid reinforcement learning (RL) for developing a task-agnostic safety filter for drones, with a focus on theoretical guarantees, practical applications, and future directions. By integrating safety constraints directly into the learning process, reach-avoid RL offers a robust and scalable framework for navigating the complexities of real-world safety scenarios. The reach-avoid safety filter, in combination with deep reinforcement learning and game-theoretic approaches, offers a feasible method for safe reinforcement learning across a range of tasks and environments in drone deployment.

As autonomous systems become increasingly deployed in real-world environments, ensuring their safety under anomalous conditions remains a critical challenge. This thesis investigates the use of large language models (LLMs) as reasoning agents for selecting appropriate fallback strategies in response to hazardous scenarios. We develop a prompting and evaluation methodology to assess an LLM agent's performance in three tasks: classifying scenarios as safe or hazardous, selecting an appropriate fallback for a hazardous observation, and predicting a set of multiple feasible fallbacks for a scenario. To improve decision quality, we apply introspective planning techniques that ground the LLM's reasoning in prior knowledge. We compare zero-shot reasoning with the introspective planning approach, demonstrating that the latter improves both accuracy and safe fallback selection. This work takes a step toward the integration of LLMs in the safety monitoring pipelines of autonomous systems.