Social Structure Matters in 3D Human-Human Interaction Generation

The paper introduces a novel framework for generating 3D human-human interactions (HHI) by addressing the modeling of social structures that dictate interaction dynamics. It presents a planner-executor paradigm, "Think with LLM, Move with Motion Skill," where a large language model (LLM) is utilized to decompose interactions into phases and assign roles, while a motion executor, enhanced with LoRA and conditioning techniques, translates this structure into coordinated motion. This approach improves the generation of physically plausible and interaction-aware 3D motions, which is critical for practitioners aiming to create realistic simulations in AI-driven environments.