Comparing AI-Assisted Authoring by Demonstration to Manual Authoring of Augmented Reality Maintenance Instructions
Year: 2026
Type: conference-paper
Abstract: Creating Augmented Reality (AR) instructions for industrial guidance remains labor-intensive, while fully automated solutions are often impractical. To bridge this gap, we present VisAARD, an open-source, human-in-the-loop authoring tool that combines vision-language models with authoring by demonstration. From stepwise expert demonstrations, VisAARD uses egocentric video and hand-tracking to generate textual descriptions and propose in-situ 3D guidance elements for later review and refinement. In a within-subject study (N=21), we compared VisAARD against an industry-standard manual authoring baseline similar to Microsoft Dynamics 365 Guides in a robotic-cell maintenance scenario. Compared to the baseline, VisAARD reduced mean authoring time from 36 to under 20 minutes, lowered perceived task load, achieved higher perceived usability, and was preferred by 20 out of 21 participants. Qualitative feedback further suggests that a hybrid workflow combining the efficiency of AI-assisted authoring with targeted manual refinement is perceived as a meaningful direction for AR authoring tools.
Source: Mensch und Computer (MuC)
Language: English
Subfield: Computer Vision and Pattern Recognition
Field: Computer Science
Domain: Physical Sciences
Open Access status: hybrid
OpenAlex ID: W7203779181