SOCIAL STRUCTURE MATTERS IN 3D HUMAN-HUMAN INTERACTION GENERATION

Traditional text-to-motion models excel at synthesizing individual human movements but struggle with the complex coordination required for human-human interaction (HHI). The authors investigate whether HHI generation can be improved by explicitly modeling the underlying social structure—the temporal phases and partner-aware roles that govern how two people interact—rather than treating it as a simple extension of solo motion generation.

The researchers first conduct a diagnostic study using a large language model (LLM) to determine if it can handle both the semantic planning and the physical execution of HHI. They find that while LLMs are excellent at "thinking" (decomposing interactions into phases like 'approach' or 'contact' and assigning roles), they fail at "moving" (generating physically plausible, continuous 3D motion).

Based on this, they propose a Planner-Executor paradigm:

1. Social Structure Planner: An LLM converts high-level text prompts into structured, motion-aligned social supervision. This involves breaking the interaction into phases and assigning specific roles to each actor.
2. Solo-to-Social (S2S) Executor: A pretrained solo motion model is adapted using Low-Rank Adaptation (LoRA). It incorporates phase-wise self-conditioning to ensure temporal continuity and ego-relative partner conditioning to ensure the actors remain spatially aware of one another.

The study demonstrates that separating social reasoning from motion synthesis significantly improves HHI generation. By using the LLM to provide explicit social structure, the model achieves better phase consistency, more accurate role alignment, and improved inter-person geometry compared to existing baselines that treat HHI as a monolithic generation task. The S2S framework effectively bridges the gap between high-level semantic intent and low-level kinematic execution.