Harnessing Coupled Stream Completion for Human-Object Interaction Modeling
Introduction
TRACE is a structured continuous-latent framework for text-conditioned Human-Object Interaction generation, stream completion, and HOI understanding.
Human-Object Interaction motion consists of body motion, object trajectories, and detailed hand articulation. These components differ in scale and dynamics but must remain coordinated in contact, relative pose, and interaction timing. TRACE separates them into body, object, and hand latent streams while coupling their dynamics through a shared latent flow.
Each stream-specific velocity is predicted from the complete evolving HOI state. Geometric constraints are applied to decoded motion to supervise contact, object-relative anchoring, and interaction timing. The same representation supports Text-to-HOI generation and completion of any single missing stream.
Frozen coupled-flow features are further connected to a language model for HOI understanding. This repository provides inference checkpoints for InterAct, OMOMO, and BEHAVE.