KITE: Decoupling Kinematics and Interaction for Zero-Shot Cross-Embodiment Manipulation

The paper presents KITE (Kinematic Interaction Transfer across Embodiments), a method for zero-shot cross-embodiment manipulation that separates task reasoning from embodiment-specific motor control. KITE utilizes a shared policy for predicting latent intents from source demonstrations and an intent-conditioned action decoder tailored to each embodiment's kinematic model, allowing for efficient adaptation to new embodiments without additional data collection. Evaluations demonstrate KITE's superior performance over existing methods in transfer success across various manipulation tasks involving different robotic embodiments.