MuseGesture: A Framework for Gesture Synthesis by Virtual Agents in VR Museum Guides

This paper presents an innovative framework named MuseGesture, designed to generate contextually adaptive gestures for virtual agents in Virtual Reality (VR) museums. The framework leverages the robust language understanding and generation capabilities of Large Language Models (LLMs) to parse tour narration texts and generate corresponding explanatory gestures. Through reinforcement learning and adversarial skill embeddings, the framework also generates guiding gestures tailored to the virtual museum environment, integrating both gesture types using conditional motion interpolation methods. Experimental results and user studies demonstrate that this approach effectively enables voice-command-controlled virtual guide gestures, offering a novel intelligent guiding system solution that enhances the interactive experience in VR museum environments.