Leveraging Gaze for Potential Error Prediction in AI-support Systems: an Exploratory Analysis of Interaction with a Simulated Robot

With the increasing prevalence of AI–driven systems that assist humans in daily tasks, ensuring safety and fostering trust is pivotal. Moreover, addressing system errors is important, as conventional methods may be inaccessible and time consuming for people with motor impairments. This study explores implicit communication using eye-tracking features. We conducted a Virtual Reality (VR) experiment where participants used a robotic arm, encountering errors to be prevented through explicit communication. Findings reveal significant differences in eye-tracking features based on error recognition and demonstrate automated error detection feasibility. Using machine learning, we were able to classify when a user intends to stop the system before a potential collision using the eye features (86 – 90% accuracy). This research contributes to understanding eye movement behavior for error detection and offers insights for implementing implicit communication in AI –- driven assistive systems.