Event

KITE: Zero-Shot Cross-Embodiment Manipulation via Decoupled Kinematics and Interaction

1 reports · 1 sources

Qianxu Wang, Shengmiao Samuel Jin, Kuan Fang, and others propose KITE, which decomposes manipulation into embodiment-agnostic interaction intent prediction and embodiment-specific kinematic decoding, connected through a contact-pattern-based interaction intent representation, to achieve zero-shot cross-embodiment manipulation policy transfer. The interaction policy is trained only on demonstration data from the source robot and is shared by target robots with different morphologies; the kinematic decoder is learned from synthetic data generated by each embodiment's kinematic model. The study evaluates the method on four manipulation tasks in simulation and real environments, covering transfer among parallel grippers, dexterous hands, and composite embodiments.