Description

Human interactive behavior prediction in human–robot collaboration is challenged by latent intention ambiguity and limited annotated interaction data. We propose a unified prediction framework that integrates implicit intention modeling, few-shot interaction learning, and conditional structured motion representation. A latent intention variable is introduced to capture interaction dynamics without explicit labels, while a few-shot training strategy improves generalization under scarce supervision. Furthermore, a conditionally structured motion embedding decouples task semantics from motion dynamics, enabling cross-task transfer.

The work is under the scope of the following projects:

• SOCIAL PIA: Cooperative Social PIA model for Cybernetics Avatars (Moonshot Research and Development Program) (web)
• TORNADO: foundaTion mOdels for Robots that haNdle smAll, soft and Deformable Objects (web)