Despite ongoing digitization in industry, many companies still work with paper instructions or ‘paper-on-glass’ solutions (e.g., PDF files on screens). In recent years, various digital work instruction (DWI) technologies have become available that provide shop-floor employees with information during their activities, e.g., sequences of instructions for tasks at hand. Engineering new instructions in these systems for new products or product variants is however expensive and time-consuming. To scale up, there is a need for methods to generate work instructions (semi) automatically. Recently, Generative AI models and Large Language Models (LLMs) have taken center stage with their abilities to interact fluently with humans, both in understanding user questions/statements and in convincingly producing natural language texts. These models however suffer from several problems, including hallucinations where unsubstantiated content is presented as facts and lack of domain-specific data about products and procedures. For instruction generation however, we need verifiably correct statements about the task at hand. To tackle both problems, we have created a pipeline that combines the generative abilities of LLMs with explicit domain-specific data. We deploy a variant of Retrieval Augmented Generation (RAG) and incorporate an ontology that augments the instructions with additional information (policies, warnings, tools). Our results show an increase in correctness of output.
An Approach to Automated Instruction Generation with Grounding Using LLMs and RAG
Lect.Notes Mechanical Engineering
European Symposium on Artificial Intelligence in Manufacturing ; 2024 ; Athens, Greece October 16, 2024 - October 16, 2024
22.03.2025
10 pages
Aufsatz/Kapitel (Buch)
Elektronische Ressource
Englisch
Visually Grounding Language Instruction for History-Dependent Manipulation
Deutsches Zentrum für Luft- und Raumfahrt (DLR) | 2022
|Automated grounding method of mobile communication vehicle
Europäisches Patentamt | 2022
|Automated Adaptive Instruction for Embedded Training
British Library Conference Proceedings | 1990
|Europäisches Patentamt | 2017
|