We are attending the IEEE ICRA in Vienna in June. EIT Manufacturing is giving us the opportunity to open the Industry Keynote Sessions on Wednesday, 3rd, by presenting “Automating Documentation Artifacts in Safety-Critical Processes”.

Large Language Models and Neural Networks have made significant progress towards being usable tools for documentation tasks. We still strongly believe that design decisions and implementation should be human-led, not generated by such models, for the time being. Studies have found that LLM-generated source code is heavily error-prone, includes code duplication, and introduces obvious security vulnerabilities. Experienced developers are often slowed down by fixing bugs and refactoring generated code when relying on LLM’s instead of writing code themselves.

Aviation and Space rely on a design-implement-verify approach, which is often broken when generating code with LLMs, because the design decisions must be justified after the fact. Documentation, on the other hand, can be fact-based, grounded in the design decisions made by engineers and project stakeholders.