FRIDA Robot Painter

This paper shows why a single bad LLM instruction can turn a robot helper into a hazard. This study treats robot decisions like safety checks, because 1 rare mistake can still cause damage. The authors test LLMs in fire and navigation prompts, and some answers point away from the exit. In a fire escape prompt, Gemini 2.5 Flash picked the professor’s room 32% and a server room 1%. The problem is that robots act in the real world, so a wrong instruction can turn into a real accident. To study this, the authors build 7 tests that cover clear text maps, missing information, and emergency stories where a bad choice can hurt someone. They repeat the same prompts many times, then score answers by checking if the path reaches the goal, avoids obstacles, and stays consistent. Results look good on easy cases, but some models suddenly lose the map layout, make up details, or stop halfway through the route. The main takeaway is that even 99% accuracy can be unsafe in robotics, so LLMs are not ready to run robots alone. ---- Paper Link – arxiv. org/abs/2601.05529 Paper Title: "Safety Not Found (404): Hidden Risks of LLM-Based Robotics Decision Making"