Human-in-the-Loop Uncertainty Analysis in Self-Adaptive Robots Using LLMs

Self-adaptive robots operating in unpredictable environments face constant uncertainty that can cause safety violations or operational failures. Traditional methods for identifying these uncertainties are ad hoc, often missing critical failure modes hidden in edge cases. Now, a team led by Hassan Sartaj has developed RoboULM, a human-in-the-loop approach that harnesses large language models (LLMs) to systematically explore and catalog uncertainties during the design stage. The tool guides practitioners through a structured prompting process, iteratively refining the analysis to uncover sources, impacts, and potential mitigations. The researchers also built a comprehensive uncertainty taxonomy tailored to self-adaptive robots, giving teams a reusable framework for risk analysis.

To validate RoboULM, the team tested it with 16 practitioners from four different industrial robotic use cases. Participants found the tool both useful and easy to understand, with high marks for its structured prompting and iterative refinement capabilities. The results suggest that LLMs, when guided by human expertise, can turn the messy task of uncertainty analysis into a systematic, repeatable practice. For engineering teams building robots that must adapt in the wild — from autonomous drones to warehouse robots — RoboULM offers a practical way to catch and address risks long before the robot hits the floor.