HOI Generation

Human-object interaction generation studies how to create plausible motion and visual sequences in which people grasp, manipulate, carry, or otherwise interact with objects. Common approaches use diffusion and flow-matching models conditioned on text, images, poses, or sparse video controls, often adding contact maps, affordances, 3D geometry, or physics-based refinement. Researchers evaluate motion and video on datasets such as GRAB, OakInk, ARCTIC, and BEHAVE, while also exploring multi-person interactions and transfer to robot manipulation. A central challenge is balancing semantic control and visual quality with accurate contact, physical plausibility, and generalization to unseen objects and scenes, especially when paired training data is scarce.