Context-Aware Multimodal Efficient Navigation in VR

Navigation is a fundamental task in virtual reality, where efficiency is crucial. Conventional virtual reality navigation methods often require users to learn complex controls, struggle to interpret user intent, and become inefficient in complex environments. We propose VR Guider, a multimodal navigation method incorporating a large language model (LLM). Our method integrates natural language, gestures, head-motion and field-of-view (FoV) cues to enable adaptive semantic navigation. Specifically, the system constructs a structured object-level scene representation and aligns it with multimodal user inputs. The LLM reasons over this context to infer user intent and generates structured navigation commands for adaptive viewpoint control. A user study in two virtual environments showed that VR Guider significantly improved task-completion efficiency relative to representative baselines and achieved the highest usability score (SUS = 87.97). A controlled ablation study further showed that multimodal cues, interaction history, and LLM reasoning improved target selection.