Human-Human-Object Interaction Generation

This field generates coordinated motion in which people interact with one another and, in human-human-object settings, with shared objects. Common approaches include text-conditioned diffusion models, variational autoencoders, and autoregressive motion models, with language models sometimes used to plan interaction phases and roles. Contact representations, collision-aware constraints, and benchmarks such as InterHuman and Inter-X help evaluate coordination, realism, and physical plausibility. A central challenge is producing diverse, physically sound interactions over long time horizons despite limited data and the difficulty of jointly modeling individual motion, social structure, and contact.