Automated LTL Specification Generation from Industrial Aerospace Requirements
This paper introduces AeroReq2LTL, an LLM-based framework that automates the translation of complex industrial aerospace natural language requirements into formal Linear Temporal Logic (LTL) specifications by utilizing a data dictionary for jargon normalization and a template-based language to clarify temporal and logical structures, achieving high precision and recall on real-world datasets.
Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer
The Big Picture: The "Lost in Translation" Problem
Imagine you are building a very expensive, life-or-death robot (like a satellite). You have a thick manual written by engineers in "Human Language" (English). It says things like, "If the sun isn't seen for 12 minutes, switch to the backup camera immediately."
Now, you need to give these instructions to the robot's brain. But the robot doesn't speak English; it only speaks a strict, mathematical code called LTL (Linear Temporal Logic). This code is like a super-precise recipe where every step must be timed perfectly.
The Problem:
In the past, humans had to manually translate the English manual into this math code. This was like trying to translate a Shakespeare play into a computer programming language while blindfolded. It took forever, was incredibly boring, and if you made one tiny mistake (like forgetting the word "immediately"), the satellite could crash.
Recently, people tried using AI (Large Language Models) to do the translation. But the AI kept failing on real satellite manuals. Why? Because the manuals are messy. They use weird jargon, hide important clues in tables, and assume you already know how the satellite works. The AI would guess the wrong variables or miss the timing, leading to a "recipe" that looked good but would poison the robot.
The Solution: AeroReq2LTL (The "Super-Translator")
The authors of this paper built a new system called AeroReq2LTL. Think of it not as a simple translator, but as a two-step "Smart Editor" that prepares the text before the AI translates it.
Here is how it works, using a Restaurant Analogy:
Step 1: The "Menu Decoder" (SpaceKG)
The Problem: In a restaurant, a waiter might say, "The chef needs the 'Red Stuff' for the special." If you don't know the kitchen, you don't know if "Red Stuff" is ketchup, chili powder, or a rare spice.
The Fix: The system has a Data Dictionary (called SpaceKG). Before the AI tries to translate, it looks at the "Interface Tables" (the kitchen's inventory list).
- It sees that "Red Stuff" in the manual actually corresponds to a specific sensor named
flagSP. - It sees that "12 minutes" corresponds to a timer variable named
deTCount. - Result: It replaces the vague "Red Stuff" with the precise "Sensor A." This stops the AI from guessing.
Step 2: The "Script Rewriter" (SpaceRDL)
The Problem: Humans are bad at being explicit. We say, "If it rains, take an umbrella." We don't say, "IF (Rain == True) THEN (Action = Take Umbrella) AT (Time = Now)." The AI gets confused by the missing "Time = Now."
The Fix: The system forces the text into a Structured Template (called SpaceRDL). It acts like a strict editor who forces the engineer to rewrite their sentence into a rigid format:
- Original: "If the sun is lost, switch modes."
- Rewritten by System: "In [Current Mode], if [Sun Lost] happens, then [Switch Mode] must happen [Immediately]."
- Result: The hidden timing clues (like "immediately") are now written in bold. The AI can no longer ignore them.
Step 3: The "Final Translator"
Now that the text has been cleaned up (variables are named correctly) and structured (timing is explicit), the AI translates it into the final math code (LTL). Because the input was so clean, the output is almost perfect.
Why This Matters (The Results)
The team tested this on a real satellite control system with 79 complex requirements.
- Old AI methods: Got about 35-50% of the translations right. They were often wrong about when things should happen or what variables to use.
- AeroReq2LTL: Got 85-88% of the translations right.
The Analogy:
If the old AI was like a tourist trying to order food in a foreign country using a broken phrasebook, AeroReq2LTL is like a tourist who first consults a local guide to learn the specific names of dishes and the cultural rules of ordering, then speaks to the waiter.
The "Secret Sauce"
The paper highlights two main innovations that make this work:
- Grounding: Connecting the words in the text to the actual code variables (like connecting "Sun" to
flagSP). - Structure: Forcing the messy human text into a clean, logical template before translation.
Conclusion
This paper shows that you can't just throw raw, messy industrial documents at an AI and expect it to work. You have to build a "bridge" first. By cleaning up the jargon and forcing the logic into a clear structure, AeroReq2LTL turns a dangerous, error-prone manual process into a reliable, automated assembly line. This means safer satellites, faster development, and fewer crashes caused by translation errors.
Drowning in papers in your field?
Get daily digests of the most novel papers matching your research keywords — with technical summaries, in your language.