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

This paper introduces RoboULM, a human-in-the-loop methodology and tool leveraging large language models to help practitioners systematically identify, analyze, and mitigate uncertainties in self-adaptive robots during the design phase, as validated by positive feedback from industrial practitioners across four use cases.

The Gist

Imagine you are building a robot that needs to navigate a busy city, fix a laptop, or sail a ship. The world is messy, unpredictable, and full of surprises. If your robot isn't prepared for these surprises (which the paper calls "uncertainties"), it might crash, break something, or get stuck.

The problem is that figuring out all the possible things that could go wrong is incredibly hard. It's like trying to list every single way a house could catch fire before you even build it. Usually, engineers have to guess based on their experience, which often misses hidden dangers.

This paper introduces a new tool called RoboULM to help solve this. Think of RoboULM as a super-smart, tireless assistant that helps engineers brainstorm every possible "what if" scenario before the robot is ever built.

Here is how it works, using simple analogies:

1. The "Master Checklist" (The Taxonomy)

First, the researchers created a giant, organized "Master Checklist" called UncerTax.

• The Analogy: Imagine a mechanic's manual that doesn't just list car parts, but categorizes every possible thing that could go wrong: Is it a flat tire (hardware)? Is it a confusing map (software)? Is it a sudden rainstorm (environment)?
• What it does: This checklist helps the robot's human engineers and the computer assistant speak the same language. It ensures they don't just think about "broken parts," but also about "confusing data" or "ethical issues."

2. The "Brainstorming Partner" (The LLM)

The tool uses a Large Language Model (LLM), which is like a very knowledgeable but sometimes chatty AI.

• The Analogy: Imagine you are trying to find a needle in a haystack. You ask a friend (the AI) to help. If you just say, "Find the needle," they might miss it. But if you give them a specific strategy, they get much better at it.
• What it does: RoboULM doesn't just ask the AI to "guess." It gives the AI a specific set of instructions (prompts) based on the Master Checklist. It tells the AI: "Look at the robot's requirements, and tell me exactly where the risks are, using these 12 specific categories."

3. The "Human-in-the-Loop" (The Refinement)

This is the most important part. The AI isn't left alone to do the work; a human is always in the driver's seat.

• The Analogy: Think of the AI as a junior intern who is eager but sometimes makes mistakes. You (the senior engineer) review their work.
  • Ranking: You give the intern a score. "You got the 'safety' part right (10/10), but your 'hardware' guess was weak (3/10). Try again."
  • Examples: You say, "Remember that time the robot slipped on a wet floor? Think about that when you guess the risks."
  • Checklist: You point to the Master Checklist and say, "You missed the 'environment' category. Go back and fill that in."
• What it does: The tool lets the human engineer keep refining the AI's answers until they are perfect. It's a back-and-forth conversation, not a one-time command.

4. The Real-World Test

The researchers tested this tool with 16 real experts who work with four different types of robots:

1. Autonomous Mobile Robots (like delivery bots in warehouses).
2. Industrial Disassembly Robots (robots that take apart laptops).
3. Collaborative Manufacturing Robots (robots working side-by-side with humans in factories).
4. Autonomous Vessels (self-driving ships).

The Results:

• The experts found the tool very useful and easy to understand.
• They loved the structured prompts (the clear instructions given to the AI).
• They found the iterative refinement (the ability to grade the AI and ask it to try again with examples) to be the most helpful part.
• The experts felt that this tool helped them find risks they might have otherwise missed, making the robots safer before they ever hit the real world.

Summary

In short, RoboULM is a digital workshop where human engineers and a smart AI work together. The human provides the experience and the final judgment, while the AI acts as a powerful engine that scans through a massive "Master Checklist" to find potential dangers. By working together in a loop of asking, checking, and refining, they can build safer, more reliable robots that are ready for the unpredictable real world.

Technical Summary

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

Problem Statement
Self-adaptive robots (SARs) operate in dynamic, unpredictable environments where unaddressed uncertainties can lead to safety violations and operational failures. While early identification of uncertainties during the design phase is more cost-effective than post-deployment remediation, systematically analyzing these uncertainties remains a significant challenge. Existing approaches often rely on intuition, prior experience, or static taxonomies, which uncover only limited subsets of uncertainties and struggle to keep pace with evolving robotic technologies and the integration of Large Language Models (LLMs). The core problem is the lack of a systematic, rigorous, and scalable method for practitioners to explore and categorize uncertainties in complex SARs at the design stage.

Methodology: RoboULM
To address this, the authors propose RoboULM, a human-in-the-loop methodology and tool designed to support practitioners in systematically exploring uncertainties using LLMs. The methodology integrates three novel components:

1. Uncertainty Taxonomy (UncerTax): A structured categorization of uncertainties derived from four industrial case studies and validated by practitioners. UncerTax organizes uncertainties across 12 dimensions: nature (static/dynamic), type (epistemic/aleatoric), stage (design/development/testing/operational), temporal duration, occurrence source (hardware/environment/software), adaptation source, scope (local/global), risk severity, affected quality attributes, propagation patterns, data characteristics, and ethical implications.
2. LLM-Driven Exploration: The tool utilizes LLMs to reason about system requirements and identify uncertainties. It employs four specific prompting strategies:
   • Role-based prompting: Establishes a persona-driven context.
   • Rubric-based prompting: Incorporates human-assigned rankings for qualitative refinement.
   • Few-shot prompting: Provides examples for experience-based guidance.
   • Ontology-constrained prompting: Guides the model using UncerTax elements.
3. Iterative Refinement Workflow: RoboULM operates via a three-step process:
   • Context Comprehension: The user provides system requirements and role definitions; the LLM summarizes its understanding of the robotic context.
   • Initial Query: The user poses an uncertainty-related question. The LLM responds with a structured output categorized according to the 12 UncerTax dimensions, providing reasoning for each.
   • Iterative Refinement: If the initial response is unsatisfactory, the user refines the output using one of three methods:
     • Ranking-based refinement: The user scores response segments (1–10) to highlight areas for improvement.
     • Example-driven refinement: The user provides concrete real-world scenarios to clarify intended interpretations.
     • Taxonomy-guided refinement: The user selects specific taxonomy elements to direct the LLM's reasoning.

The tool is implemented as a web application (React frontend, Express backend) using the Gemini 2.5 Flash model (chosen for its hybrid reasoning and large context window), though it is compatible with other models like ChatGPT and Llama.

Key Contributions

• UncerTax: A comprehensive uncertainty taxonomy specifically for SARs, detailing identification methods, sources, impacts, and mitigation strategies across 12 dimensions.
• RoboULM Tool: A functional prototype that operationalizes human-in-the-loop uncertainty analysis, combining structured prompting with iterative refinement capabilities.
• Prompting Strategies: A suite of advanced prompting techniques (ranking, example-driven, and taxonomy-guided) designed to iteratively refine LLM-generated uncertainty analyses.

Evaluation and Results
The authors evaluated RoboULM with 16 practitioners (including researchers, software engineers, and robotic engineers) across four industrial use cases: Autonomous Mobile Robot (AMR), Industrial Disassembly Robot (IDR), Collaborative Manufacturing Robot (CMR), and Autonomous Vessel (AV).

• Usability: The tool was generally perceived as useful and easy to understand. The AMR case study received the highest ratings (median 4.5/5 for utility and understanding).
• Feature Preference: Structured prompting was the most consistently valued feature (Mean: 4.25, Top-Two-Box score: 87.5%). Iterative refinement (Step 3) was identified as the most helpful aspect by participants.
• Refinement Methods: Ranking-based refinement was the most frequently used and perceived as straightforward. Example-driven refinement received strong ratings when participants could provide relevant examples. Taxonomy-guided refinement was less favored by some participants (particularly in CMR and AV cases) due to the challenge of identifying relevant taxonomy elements and occasional LLM focus on explaining the taxonomy rather than refining the response.
• Interaction Logs: Most participants completed the tasks, exploring multiple uncertainty questions. The analysis confirmed that participants covered a broad range of uncertainty sources (environment, hardware, software, human) and impacts.

Significance and Claims
The paper claims that RoboULM demonstrates the practicality of using LLMs for systematic uncertainty analysis in complex robots. The findings suggest that combining structured prompting with human-guided iterative refinement allows practitioners to explore uncertainties more comprehensively than traditional methods. The authors position RoboULM as a viable solution for addressing the challenge of rigorous uncertainty analysis in SARs, particularly by leveraging the reasoning capabilities of LLMs while maintaining human oversight through the taxonomy and refinement loops.

The authors remain modest regarding their claims, noting that this work presents a "first working prototype" and a usability study. They acknowledge limitations in external validity due to the sample size (16 participants) but argue that the diversity of industrial cases and participant roles strengthens the findings. Future work is planned to incorporate participant-suggested features and evaluate the tool across additional robotic cases and LLMs.