This research introduces a novel approach combining multi-modal large language models with Gaussian splatting for efficient physics simulations in 3D scenes without manual property assignment.

• Creates physically plausible simulations by using MLLMs to extract object properties from images
• Achieves 40× speedup compared to traditional physics engines while maintaining simulation quality
• Enables customizable physics interactions through natural language instructions
• Supports diverse object types and complex interactions in 3D environments

This advancement significantly reduces the engineering effort needed for realistic physics simulations in gaming, AR/VR, and digital twins applications by automating property assignment and optimizing computational requirements.

Efficient Physics Simulation for 3D Scenes via MLLM-Guided Gaussian Splatting