Exploring Large Language Models for Digital Human Modeling

Abstract

Digital Human Modeling (DHM) tools, such as IPS IMMA, are used to evaluate industrial ergonomics and assembly feasibility before physical implementation. However, their reliance on action-centric graphical user interfaces can create steep learning curves and require repeated graphical user interface operations, such as scene setup, manikin placement, object manipulation, and ergonomic analysis. Recent advances in Large Language Models (LLMs) create opportunities for intent-centric Natural Language Interfaces (NLIs), where users describe tasks naturally rather than performing low-level operations. This thesis develops a locally deployed, LLMbased NLI that translates high-level intents into validated IPS IMMA commands, and compares its usability with the traditional GUI. The proposed system uses a hybrid three-tier architecture connecting a local LLM to the simulation environment via its scripting interface. LLM-based intent interpretation is combined with rule-based validation, session context, and pipelinebased inference to resolve missing parameters and generate structured commands before execution. By constraining LLM outputs before execution, the system is designed to reduce the likelihood that hallucinated or unsupported commands are passed to the simulation engine. A geometry-based spatial reference method is also implemented to map natural-language directions to stable coordinate frames in the simulation environment. A within-subjects comparative study with 10 participants, in which each participant used both the NLI and the graphical user interface, was conducted using the ISO 9241-11 usability framework, which defines usability in terms of effectiveness, efficiency, and satisfaction in a specified context of use. The evaluation combined task performance metrics, System Usability Scale (SUS) scores, and open-ended questionnaire responses to assess effectiveness, efficiency, and satisfaction. The evaluation results show that the proposed NLI supported the selected IPS IMMA workflow in the tested scenario. Compared with the GUI, the NLI condition achieved a higher task completion rate and higher perceived usability scores, shifting interaction effort from mouse-based operations toward typed input. Qualitative feedback suggested that the NLI was highly useful for broad task commands and initial setup, whereas the GUI remained important for fine-grained adjustments and troubleshooting. The findings indicate that locally deployed LLM-based NLIs have strong potential as complementary interfaces for DHM workflows when combined with constrained action spaces, validation mechanisms, and clear user feedback.