Imagine a delivery drone smoothly navigating a busy city. Suddenly, it spots something unexpected—a flock of birds, a sudden gust of wind, or an unexpected obstacle. How does it react in real-time to avoid a collision? Researchers are exploring how large language models (LLMs) can empower robots to handle these unforeseen events. LLMs, trained on vast amounts of text and code, can reason in a way that mimics human intuition. In this new research, a two-step system called AESOP has been designed to give robots a fast and a slow kind of thinking, much like humans. The "fast" thinking uses embeddings, which are like compressed representations of past experiences. When the robot sees something new, it quickly compares it to these embeddings to detect if it's an anomaly. If an anomaly is detected, the "slow" thinking kicks in. This part uses the full power of the LLM to reason about the potential danger. For instance, a bird far off in the sky might be inconsequential, but a flock of birds directly in the drone's path requires immediate action. AESOP integrates this two-stage reasoning with a clever control strategy. The robot continually plans multiple escape routes, ensuring it can always execute a safe maneuver, even while waiting for the LLM to finish "thinking." Tests on simulated drones and even real hardware show promising results. Drones were able to avoid collisions by switching to pre-planned safe landing zones or holding patterns when faced with anomalies. What's exciting is that the fast anomaly detection doesn't need a massive LLM, making it suitable for robots with limited onboard computing power. This research is a big step towards more reliable and safer robots in the real world. Future research might explore how to make the LLM's reasoning even faster or train robots to learn from their experiences with anomalies. The ability for robots to react to unexpected events is crucial for their widespread adoption, and this research offers a compelling approach to tackling this challenge.