Reproduce “HOI-Dyn: Learning Interaction Dynamics for Human-Object Motion Diffusion” on OpenResearch

About this paper

Lin Wu, Zhixiang Chen, Jianglin Lan

Researchers introduce HOI-Dyn, a framework that models 3D human-object interaction generation as a driver-responder system, where human actions causally drive object motions. This approach, which incorporates a lightweight interaction dynamics model and a residual-based auxiliary loss, yields more physically plausible and causally consistent interactions, outperforming prior methods on metrics like human motion quality and interaction fidelity.

Key insights

• Framing human-object interaction as a driver-responder system inherently promotes causal consistency and physical plausibility, with contact dynamics emerging naturally from the interaction rather than being explicitly modeled.
• A lightweight, auxiliary dynamics model can effectively guide a generative diffusion process, even with its own prediction errors, through a residual loss formulation that targets differential consistency.
• The interaction dynamics loss can serve not only as a training objective but also as a quantitative metric for evaluating the causal consistency and overall quality of generated human-object interactions.

Summary by alphaXiv.