Interactive Art
A New Model Figures Out How a Virtual Body Should Touch an Entire Scene, Not Just One Prop
UniffordInteractor, published in ACM Transactions on Graphics as part of SIGGRAPH Asia 2026, tackles a gap in human-object interaction generation: prior methods handle a single object well but struggle once a scene has several objects that also contact each other. Its two-stage 'reason-then-interact' framework uses a fine-tuned 3D multimodal LLM to predict object-centric affordances, then a hierarchical generator that produces motion autoregressively from root to body to hand to object, conditioned on that reasoning.