KITE: Decoupling Kinematics and Interaction for Zero-Shot Cross-Embodiment Manipulation
KITE decouples manipulation into two parts: predicting general interaction intent and controlling specific robot movements. It uses a shared policy trained on one robot's demonstrations to predict an embodiment-agnostic intent, which is then decoded into actions for any new robot using only its kinematic model. This allows skills learned on one physical robot to be transferred zero-shot to structurally different robots without needing new task demonstrations.