Back to Research

Balancing Initiative in Embodied Navigation: Comparing Conventional, Automatic, and Proactive AI Guidance in a VR Kitchen Task

This study investigates how three AI assistance strategies—Conventional (user-initiated), Automatic (time-based), and Proactive (behavior-triggered)—shape wayfinding performance and user experience in a VR kitchen task. Using behavioral logs, trajectories, and event data from three participants across all conditions, we observe consistent patterns: the Conventional assistant leads to prolonged hesitation and delayed help-seeking; the Automatic assistant reduces disorientation but disrupts natural exploration through frequent, time-driven prompts; and the Proactive assistant intervenes only when behavioral cues indicate genuine confusion, providing timely support without interrupting ongoing activity. These results offer preliminary evidence that Proactive assistance achieves the best balance between efficiency and experiential quality—improving navigation performance while preserving user agency and reducing cognitive load. We discuss design implications for embodied AI guidance, including respecting bodily cues, aligning intervention timing with user rhythms, and avoiding over-automation in XR navigation systems.

By defining explicit role instructions and preloading the spatial layout of objects in the environment, the LLM-based agent gains both spatial understanding and contextual knowledge, enabling it to more effectively support user interaction and object-finding tasks.

An assistant that automatically provides guidance whenever the user remains idle or makes no meaningful action for more than 15 seconds.

A help-on-demand assistant that only appears when the user actively requests guidance.

A behavior-aware assistant triggered by a hidden human observer when clear signs of confusion—such as standing still, turning in circles, or backtracking—are detected.

实验数据文件下载